You are on page 1of 189

Gianluca Valentini

PHOTONICS
Part II
Notes for the MS Students in Engineering Physics
at Politecnico di Milano
Ver. 2.4

Politecnico di Milano 1



Table of content

Fundaments of Radiometr 5
Introductio 6
Basic concepts of geometrical optic 8
Thin lenses and optical system
Stops of an optical syste 1
Aperture stop, Entrance Pupil and Exit Pupil 13

Field Stop, entrance window and exit window 15

Aplanatic optical system 1


The Lagrange theorem 17

The Abbe’s sine condition 18

The radiometric quantitie 20


Law of conservation of the radianc 23
Free space propagatio 2
Interface between two optical medi 2
Lambertian source 28
Irradiance produced by extended Lambertian source 3
Lambertian disk 31

Lambertian sphere 34

Integrating sphere 35

Radiometric properties of image 37


Radiance of an imag 3
Irradiance of the image of a lambertian sourc 3
Introduction to photometr 43
Anatomy and physiology of the human eye 43

Simpli ed description of the visual process 45

The Natural light and the human eye 47

Conclusive remark 50
Exercise 52
Noise in Photodetector 56
Introductio 57
Thermal noise or Johnson nois 63
Shot Nois 66
Photodetector 72
Introductio 73
Basic concepts of radiation detectio 75
Figures of merit for radiation detector 7
Responsivity 77

2
fi
s
e
n
n
n
s
s
s
e
n
s
m
s
s
y
y
s
e
a
s
s
e
s
n
e
s
8

Frequency response 77

Noise equivalent power 78

Thermal detector 82
Thermoelectric Detector 8
Bolometer 8
Photoacoustic detector 9
The Golay cell 91

Pyroelectric detector 9
Quantum detector 98
Photoemissive detector 9
Photoelectric e ect 99

Photocathode materials 100

Vacuum photodiodes 103

Photomultiplier tubes 108

Hybrid photomultipliers 113

Noise in photomultiplier tubes 114

Streak camera 119

Photoconductive detector 12
Junction photodetector 12
pn photodiodes 126

Avalanche photodiodes 134

Noise in photodiodes 137

Homodyne detection 140

Pixelated detector 144


Introductio 145
The MOS Capacito 146
The ideal MOS capacito 14
The accumulation regim 14
The depletion regim 15
The inversion regim 15
The deep depletion regim 15
The capacitance of the MOS capacito 15
Approximate Analysis of the MOS Capacito 156
Equilibriu 15
Non Equilibriu 15
Charge Coupled Device 163
The charge storage paradig 16
Transfer of Charge between MOS Gate 16
Charge Con nement beneath the MOS Gat 16
The Output Stage of a CC 17
Realistic schemes commonly used in CCD 17
Four Phase clock CCD 172

3
m
s
n
fi
ff
m
s
s
e
e
r
s
s
s
s
s
s
e
r
s
s
e
D
m
r
s
s
e
r
4

Two Phase clock CCD 172

Buried-Channel CCD 173

The CCD image senso 17


Spectral sensitivity 176

Dark current 178

Noise in CCDs 179

Dynamic range 181

Ampli ed CCD Detectors 181

Color CCDs 183

CMOS vs. CCD image sensor 185

4
fi
r
s
4

Chapter 1

FUNDAMENTS OF RADIOMETRY
“Light brings us the news of the Universe”
- Sir William Henry Bragg


Section 1

INTRODUCTION
Light is such a common experience that everybody might believe he knows everything about it.
Nevertheless, some misconceptions about the way we see the external world through our eyes are
rather widespread amongst people. In fact, it is commonly accepted that we “see” the intensity of
light. Yet, the inconsistency of this awed belief can be revealed by simple questions:
• Why a luminous object does not become dark when we move away from it, since the amount of
light captured by our eyes decreases with the distance?
• Why the brightness of some objects keeps the same when observed from di erent directions, while
other objects change their appearance?
Radiometry and Photometry deal with the physical laws governing the propagation of energy through
radiation elds and the perception of visible light by human beings, respectively, and provide the an-
swers to the above puzzles.
Coming back to the rst question, it is known that the intensity of the electromagnetic radiation
originating from a small object, or from a large object far away from the observer, attenuates with the
square of the distance. This is nothing more than a consequence of the law of conservation of energy.
To be more precise: the transport of energy by an electromagnetic eld is described by the Poynting
vector P(r;t), which is the cross product of the electric eld E times the magnetic eld H:

P = E × H. (1.1.1)

In an electromagnetic wave, the vectors E and H are orthogonal to each other and their amplitudes
are related by:

ϵ0 1
|H| = |E| c= , (1.1.2)
μ0 ϵ0 μ0

where c is the speed of light and ε0 is the vacuum permittivity.


Because of the extremely high frequency of light, the time average of the Poynting vector over a
few periods should be conveniently used instead of P(r;t). The intensity I(r) of an electromagnetic
wave at position Q(r) in space, for the Poynting theorem, is given by the ux of ⟨P(r)⟩ across a unit
area orthogonal to P (Figure 1.1).

⟨P(r)⟩ ⋅ u n dS ⟨ | P(r) | ⟩ u r ⋅ u n dS
I(r) = = = ⟨ | P(r) | ⟩ (1.1.3)
u r ⋅ u n dS u r ⋅ u n dS

Therefore, from equations (1.1.1), (1.1.2), and (1.1.3), the intensity I(r) in Q(r) is:

I(r) = c ϵ0 ⟨ | E2(r) | ⟩ = c ϵ0 ERMS


2
(r), (1.1.4)

where ERMS is the root mean square value of the electric eld.

6
fi

fi

fl

fi
fi

fi
fl
ff
fi

The power owing across a spherical surface Σ of radius d, concentric with a small source S emit-
ting isotropically, is then given by:

∫Σ
2
Po wer = ⟨P(r)⟩ ⋅ u n d S = [c ϵ0 ERMS (d )] [4 π d 2 ] (1.1.5)

and must be constant independently on d, for the conservation of energy.


For the symmetry of the problem, the root mean square value of the electric eld ERMS is uniform
all over Σ and given by:

Po wer 1
ERMS (d ) = , (1.1.6)
4 π c ϵ0 d

so that, the intensity I(d) is proportional to 1/d2, i.e decreases with the square of the distance, as ex-
pected.

Figure 1.1 Intensity of a light source

Therefore, the amount of light entering the pupil of the eye at distance d from an isotropic source,
like a light bulb, decreases accordingly. Nevertheless, the appearance of any light source, or more
speci cally its brightness, does not change with the distance. Based on this experimental “observa-
tion” we can conclude that our eyes do not measure the light intensity, as it is commonly assumed,
but a more fundamental quantity called Luminance, which is a pillar of Photometry, as much as the
Radiance, which is its physical counterpart, is the fundamental quantity of Radiometry.
This chapter is intended to shed light on radiometric quantities and concepts, which must be taken
into account when considering the propagation of a radiation eld through free space or optical me-
dia.

Figure 1.2 Viewing a distant object

fi

fl


fi

fi

Section 2

BASIC CONCEPTS OF GEOMETRICAL OPTICS

Thin lenses and optical systems


In radiometry, light propagation is supposed to obey the laws of ray optics. This excludes dif-
fraction, which can be neglected for the purpose of studying the transport of energy by a radiation
eld through apertures much greater than the wavelength. Therefore, before discussing the main top-
ics of this chapter, it is convenient to spend some time to recall the basic concepts of geometrical op-
tics.

Figure 1.3 The Snell law.

It is known that, when an optical ray impinges on the interface between two media with di erent
refractive indices n1 and n2, the ray is refracted according to the Snell’s law:

n1 si n(θ1) = n 2 si n(θ2 ). (1.2.1)

Two spherical surfaces close to each other, delimiting a medium with refractive index di erent
from the one of the outside medium make a spherical lens. A lens is assumed to be thin when its
thickness is much lower than the curvature radii of the two surfaces. Assuming that the external
medium on both sides of the lens is air, for which n ≈ 1, a thin lens is characterised by a focal length
f, which is given by the lensmaker’s formula:

[ r1 r2 ]
1 1 1
= (n − 1) + , (1.2.2)
f

where r1 and r2 are the curvature radii of the refractive surfaces and n is the refractive index of the
medium therein. Unfortunately, di erent sign conventions are commonly used in optics. The sign
convention used to write Eq. (1.2.2) is shown in Figure 1.4. The focal length f is, by de nition, the
distance from the lens where a bundle of parallel rays is focused to a point called focal point. This
holds for a positive lens; in case of a negative lens, the prolongations of rays meet each other in the

8
fi



ff

fi
ff
ff

Figure 1.4 Sign convention for the the Lensmaker's equation.

Figure 1.5 Image of an object made by a thin lens.

focal point. For a thin lens in air, two focal point F are symmetrically located on both sides of the lens
in two planes at distance f from the lens, called front and back (or rear) focal planes (Figure 1.5).
The most familiar property of a lens with positive focal length is its ability to form a real image of
an object. Taking advantage of the cylindrical symmetry, most optical systems can be analysed consid-
ering only meridional rays, i.e. rays which lay in the plane containing the optical axis and the object
point from which each ray originates. With reference to Figure 1.5, a thin lens of focal length f makes
a magni ed replica of an object h1 placed in a plane at distance s1 from the lens (s1 > f), in a second
plane at distance s2. The replica has length h2. According to the most used convention, s1 and s2 are
taken both positive when to object is on the left side of the lens and the image is on the right side.
The distances s1 and s2 are related to the focal length f by the Lens law:

1 1 1
+ = . (1.2.3)
s1 s2 f

The planes s1 and s2 are said conjugated planes.


When the distance s1 of the object from the lens is lower than f, the lens makes a virtual image.
This means that the rays emitted by any point of the object after passing the lens diverge, while their
prolongation to the left of the lens converge into an image (virtual) that does not bring any energy,
but can be considered the “object” for a second lens (or any other optical element) that makes a com-
posite optical system.
From simple geometrical considerations referring to triangles that are similar in Figure 1.5, it is
evident that the Lateral Magni cation M of the object, also called Transverse Magni cation or
simply Magni cation Ratio, is:

| h2 | s
M=− = − 2. (1.2.4)
| h1 | s1

9
fi
fi

fi

fi

Figure 1.6 Focal distances where parallel rays are focused by


a thin lens in media with different refracting indices.

For what has been said before, the lateral magni cation is negative (M < 0) for a real image made
by a thin lens and positive (M > 0) for a virtual image.
Even if equation (1.2.3) is the one most commonly used to nd the conjugated planes of a thin
lens having focal length f, in some cases, the Newton’s lens formula can be more convenient:

x1 x 2 = − f 2 [x1 < 0, x 2 > 0], (1.2.5)

where the distances x1 and x2 are taken from the front focal plane to the object plane and from the
back focal plane to the image plane, respectively, as shown in Figure 1.5. Using the Newton’s formal-
ism, the lateral magni cation is:

| h2 | f x
M=− = =− 2 [x1 < 0, x 2 > 0]. (1.2.6)
| h1 | x1 f

Figure 1.7 Thick lens (a) and generic optical system


(b).

10
fi


fi

fi

Figure 1.8 Focal panes and principal planes of a two-lenses optical


system.

Figure 1.9 Image construction for the two lens system.

Eq. (1.2.6) tells us that moving the object closer to the focal plane leads to a larger magni cation,
which becomes in nite when the object reaches the focal plane. Accordingly, Eq. (1.2.5) shows that
the image forms at a greater distance from the lens.
Sometimes, it is necessary to consider optical systems operating in media with di erent refractive
indices. This is the case for some microscope objectives that might be immersed in water or oil on
the object side and have air on the image side. In general, it is possible to demonstrate that any opti-
cal system can be characterised by a front and back focal lengths f1 and f2, which are related to each
other and to the refractive indices n1 and n2 of the media the system is in, by the equation:

f1 f
= 2. (1.2.7)
n1 n2

Figure 1.6 gives a pictorial view of the meaning of the focal lengths f1 and f2 with reference to two
bundles of parallel rays coming form the left (red) and right (green) sides.
The simplest case of an optical system di erent from a thin lens is a thick lens, depicted in Figure
1.7(a). Note that in this case the focal lengths f1 and f2 have to be taken from two planes H1 and H2

11

fi
ff

ff

fi
called Principal Planes, where the rays originating from the focal points are conventionally bent and
run parallel to the optical axis. By de nition, the principal planes are conjugated planes characterised
by a lateral magni cation M = 1. We will come back to this point later on.
Figure 1.7(b) shows the schematic of a generic optical system made by whatsoever number of ele-
ments. The system can still be represented by a front and back focal lengths f1 and f2, being the dis-
tances measured from the principal planes. Provided that also the distances s1 and s2 are measured
from the principal planes, the positions of the conjugated planes of a generic optical system are given
by the equation:

n1 n 2 n n
+ = 1 = 2, (1.2.8)
s1 s2 f1 f2

which generalises the lens law formula (1.2.3) in case the rst and last medium have di erent refrac-
tive indices. When the refractive indices n1 and n2 of the rst and last medium are the same, usually
set to 1 in case of air, Eq(1.2.8) degenerates into Eq(1.2.3).
The principal planes H1 and H2 of a systems made by a number of optical elements are, by de ni-
tion, those conjugated planes between which the magni cation ratio is equal to +1. As an example,
consider the simple optical system shown in Figure 1.8, which is made by two thin lenses of equal
focal length f placed at a distance f from one another. The system has an equivalent focal length f,
with the front focal plane in the plane of the rst lens, and the back focal plane coincident with the
second lens. The focal planes can be easily found by considering the positions where a bundle of par-
allel rays coming either from the front or the back sides are focused to a point. Vice-versa, the rst
principal plane (H1) is in the plane of the second lens, while the second principal plane (H2) is in the
plane of the rst lens (see below the explanation). It turns out that focal planes and principal planes
are superimposed to each other, yet the indices 1,2 are exchanged, so that: F1 ≡ H2 and F2 ≡ H1. For the
speci c imaging condition shown in Figure 1.8, the lens law (1.2.3) is satis ed for the distances s1 =
2f and s2 = 2f, being s1 and s2 measured from the principal planes. In this case M = -1. You might also
see that the Newton’s formula is easily satis ed for x1 = -f and x2 = f, where x1 and x2 have to be tak-
en from focal planes F1 and F2, respectively.
It is worth noting that in the Newton’s formalism, the position of the principal planes h1 and h2,
with respect to the focal planes, assume a very simple expression, which can be derived from Eq.
(1.2.6) assuming M = +1 and solving for x1 and x2:

h1 ≜ x1 ⇒ M = 1 → h1 = f h2 ≜ x2 ⇒ M = 1 → h 2 = − f. (1.2.9)

This allows us to immediately nd the planes H1 and H2, considering that the distances h1 and h2
must be taken from the focal planes.
The general construction that can be used to nd the point P’ conjugated to a point P for whatso-
ever optical system is shown in Figure 1.7(b). Two rays have been drawn from the object point P:
one parallel to the optical axis, the other crossing the rst focal point F1. The image point P’ is at the
intersection of the two rays, which must run parallel to the optical axis between the two principal
planes.
The same construction is implemented in Figure 1.9 for the system depicted in Figure 1.8. The
rst ray (green) goes parallel to the optical axis to the principal plane H1, continues (back) to the
principal plane H2, then passes through the focal point F2. The second ray (red) passes through the
focal point F1, reaches the rst principal plane H1 and goes parallel to the optical axis to H2. The con-
jugated point P’ is at the intersection between the two rays or their prolongation, as it is the case in
f
the f ↔ f optical system. Once the principal planes are known, this construction can be done ignor-
ing the real positions of the two lenses. The same result could be achieved also applying the lens law
(1.2.3) separately for the two optical elements, which is what has been done in gure Figure 1.8.

12
fi
fi

fi

fi

fi
fi

fi

fi
fi
fi

fi
fi
fi
fi

fi
fi
ff

fi
fi
Nevertheless, the construction shown in Figure 1.9 has the advantage of providing a pictorial view to
the property of the principal planes. In fact, the two rays red and green run parallel to the optical axis
from H1 to H2, demonstrating the unitary magni cation ratio between these planes.
Finally, it is important to consider that not every light ray emitted by an object and directed toward
an optical system contributes to the formation of the nal image. Actually, many rays are blocked by
the limiting apertures made by the elements of the system (lenses) or by any diaphragm or iris inten-
tionally inserted into it. In fact, it is known that any optical device, like a photographic lens, a binocu-
lar or a microscope, has a limited Field of View and an aperture, sometimes called Numerical Aper-
ture (N.A.) or f-number (f/#), that measures the light gathering power of the system, i.e. how lumi-
nous would be the image for a given object.
Figure 1.10 shows the stops that limit the bunch of optical rays that contribute to the formation
of the image in an optical system made by two thin lenses of equal focal length f, but di erent size.
For the speci c choice of the object plane located at distance 2f from the rst lens, the system has a
magni cation M =+1. Hence, the principal planes of the system are exactly those of the object and
the image. While this is certainly a very special case, the considerations about the stops shown in
Figure 1.10 and discussed in the next subsection hold in general.

Figure 1.10 Optical system made by two lenses of equal focal lengths.
Aperture and eld stops are also shown (modi ed from “R.W. Boyd - Ra-
diometry and the Detection of Optical radiation”)

➡ Aperture stop: limits the cone angle of the ray bundle from the object to the image
➡ Field stop: limits the eld of view in the object plane
➡ Entrance pupil: is the image of the aperture stop in the object space
➡ Exit pupil: is the image of the aperture stop in the image space
➡ Entrance window: is the image of the eld stop in the object space
➡ Exit window: is the image of the eld stop in the image space

Stops of an optical system

APERTURE STOP, ENTRANCE PUPIL AND EXIT PUPIL

13
fi
fi
fi
fi
fi
fi
fi


fi

fi

fi

ff
The Aperture Stop limits the amount of light that contributes to the image. Even if the aperture
stop can be anywhere in an optical system, in many cases, an appropriate design suggests that the
ultimate aperture stop is set by the physical size of the rst lens, which is usually demanded to cap-
ture as much light as possible toward the image. Indeed, in the example shown in Figure 1.11 the
rst lens limits the cone angle of rays leaving an objet point on axis and entering the system. Hence,
the rst lens determines the extreme rays that can be accepted. One of those rays, said Marginal Ray
is highlighted in red in Figure 1.11.

Figure 1.11 Pencil of rays entering the aperture stop of an opti-


cal system. The marginal ray is highlighted in red

In many optical systems, like photographic lenses, an iris located inside the objective provides a
variable aperture, which becomes the actual aperture stop, controlling the amount of light that reach-
es the detector. The image of the aperture stop in the object space is called Entrance Pupil and limits
the portion of the rst lens that can accept light. In the optical system shown in Figure 1.10 and in
Figure 1.11, which lacks an internal iris, the aperture stop and the entrance pupil coincide.
Figure 1.12 shows the general procedure required to nd the aperture stop of any optical system:
a) the image of each physical stop (not shown in Figure 1.12) must be found in the part of the sys-
tem that precedes it;

b) the image that subtends the smallest angle at the axial point P0 located in the object plane is the
entrance pupil (red lines in Figure 1.12);
c) the physical stop whose the entrance pupil is the image, is the aperture stop (not shown here for
generality).

Figure 1.12 Windows and pupils of a generic optical system.

14
fi
fi

fi

fi
fi

In same cases, the entrance pupil and the aperture stop can coincide, as we have seen before. The
angle 2θ0 which the diameter of the entrance pupil subtends at P0 is called the Angular Aperture on
the Object Side, or simply the Angular Aperture of the system.
The image of the aperture stop in the image space is called Exit Pupil (red lines in Figure 1.12)
and is the virtual aperture through which the light has to pass to reach the detector. The angle 2θ1
which the diameter of the exit pupil subtends at the image point P1 is called the Angular Aperture
on the Image Side (sometimes also called the Projection Angle). In Figure 1.11 the exit pupil is
shown in pale red, with the marginal ray just touching it.
To better understand the importance of the exit pupil, you might consider Figure 1.13 that
schematically shows the optical rays in a microscope. The microscope is made by an objective (the
rst lens), which makes a real image of the object magni ed by a factor M1, and by an ocular (the
second lens), which further magni es the real image by a factor M2 and produces a virtual image to
be observed by naked eye. In this case, the exit pupil (green in Figure 1.13) is the circle where the
rays cross each other while exiting the optical system. Its meaning is crystal clear: the exit pupil is the
position where the “pupil” of the eye has to be put in order to observe the magni ed image of the
sample placed in the object plane. In a proper design, the size of the exit pupil matches the size of the
widely open pupil of the human eye (about 8 mm). Actually, in this case the aperture stop is not the
rst lens, but a telecentric diaphragm put in the back focal plane of the objective (light blue arrows)
to grant the same e ective aperture all over the eld of view of the microscope.

Figure 1.13 Optical rays in a microscope made by an objective lens and


an ocular

FIELD STOP, ENTRANCE WINDOW AND EXIT WINDOW

The Field Stop restrains the portion of the object plane that could be imaged by the optical sys-
tem. It may be made by a physical diaphragm, which restricts the image to that part of the object that
can be accurately observed. In some cases, this is required to remove the peripheral region of the
Field of View, which might be a ected by aberrations and severe loss of light. For example, the black
circle observable by looking through a binocular or a microscope is the image of the eld stop, as it is
seen by our eye. This is a rather general condition that holds for almost all the devices designed for
observation with the naked eye. In other optical systems, like photographic or scienti c cameras, no
such a physical diaphragm is present. Yet, in this case, the detector (e.g. a CMOS sensor) intrinsically
provides a eld stop by limiting the image size. Considering the optical system made by the two lens-
es shown in Figure 1.10, the eld stop is set by the edge of the second lens, which is the only aper-

15
fi
fi
fi
ff
fi
ff
fi
fi

fi

fi
fi
fi
Figure 1.14 Optical rays emitted by the edge of the eld of view of the
optical system shown in Figure 1.6. The chief chief ray is highlighted in
blue. The angles φ0 and φ1 represents the eld angle and the image eld
angle, respectively.

ture beyond the entrance pupil. In Figure 1.14, the Chief Ray (highlighted in dark blue), also called
Principal Ray, which is drawn from the edge of the eld stop (second lens) to the center of the en-
trance pupil ( rst lens), de nes the size of the eld of view (black arrow) in the object plane. The
chief ray also passes through the center of the exit pupil. The image of the eld stop in the object
space is called Entrance Window (shown in light blue in Figure 1.14). In general, to determine the
eld stop, as we have done for the aperture stop, we nd rst the image of each stop in the part of
the system which precedes it. That image which subtends the smallest angle (2φ0 in Figure 1.12) at
the centre of the entrance pupil is the entrance window, and the angle 2φ0 is called the Field Angle
or the Angular Field of View. The physical stop which corresponds to the entrance window is then
the required eld stop. The eld angle is twice the angle φ0 made by the chief ray to the optical axis
(Figure 1.14). The entrance window limits the portion of the object space that can be observed
though the system. The image of the eld stop in the image space is called Exit Window. The angle
(2φ1 in Figure 1.14), which the diameter of the exit window subtends at the centre of the exit pupil,
is sometimes called the Image Field Angle. In Figure 1.10 and Figure 1.14 the eld stop and the
exit window coincide and the image eld angle is twice the angle φ1 made by the chief ray to the opti-
cal axis in the image space.
In the case of the system in Figure 1.14, the eld of view is not limited by a sharp circle in the
image plane, as it happens in microscopes. Nevertheless, for the points in the object plane beyond the
tip of the arrow, the image gets very dark, since the size of the pencil of rays that can enter the sys-
tem drops abruptly, not to mention the aberrations usually su ered by the peripheral portion of the
image. Therefore, even if the size of the eld of view is conventionally de ned by the procedure de-
scribed with reference to Figure 1.12, as a matter of fact, the actual eld of view can be set by the
detector put in the image plane. The detector introduces a sharp stop that coincides with the exit
window of the system. In a properly designed imaging device the detector must have a diagonal that
subtends an angle at the exit pupil smaller than the image eld angle determined with the procedure
of Figure 1.12. More speci cally, to limit the loss of light in the peripheral part of the image, which
is said Vignetting, the diagonal of the detector should be smaller than the size of the eld of view in
the image plane, de ned by twice the length of the right arrow in Figure 1.14. Vignetting is almost
unavoidable in optical systems, even if it is usually negligible in telescopes and other systems with a
small eld angle.

16
fi
fi
fi
fi

fi

fi
fi
fi
fi
fi
fi
fi
fi
fi
fi
fi
fi
fi
fi
ff
fi
fi
fi
fi
fi
fi
Aplanatic optical systems

THE LAGRANGE THEOREM

In paraxial approximation (Figure 1.15), optical rays form a small angle with the optical axis (θ, u,
θ’, u’ << 1). Images are assumed to be exact replica of objects since aberrations (mostly spherical
aberration) are negligible.
Consider a single refracting surface, as shown in Figure 1.15. Let P, P’ be a pair of axial conjugat-
ed points, their distances from the surface being l and l’, respectively. We now place a small object at
P and draw a paraxial ray from the top of the object to the vertex of the surface Q. The ray will be re-
fracted at Q, the slope angles u and u’ being the angles of incidence and refraction. Hence n u = n′u′,
because in paraxial approximation we can assume sin(x) ≈ x. Therefore, n h /l = n′h′ /l′. Multiplying
both sides of the equation by the height y of a ray drawn from P (red arrow) gives:

n h θ = n′h′θ′ (1.2.10)

This important relation is called the Lagrange theorem.


Because h’, n’, and θ’ on the right-hand side of one surface are, respectively, equal to the same
quantities on the left-hand side of the next surface, it is clear that the product n h θ is invariant for
all the spaces between surfaces, including the object space and the image space. This product is called
the Lagrange Invariant or, more often, the Optical Invariant.

Figure 1.15 The Lagrange invariant (modi ed from “R. Kingslake - Lens design fundamentals”)

Since the Lagrange theorem applies to the object and also to the nal image, the lateral magni ca-
tion M is given by:

h′ nθ
|M | = = (1.2.11)
h n′θ′
For n = n’, the optical invariant becomes:

h′θ′ = h θ (1.2.12)

For an optical system in air, made by any number of lenses, one gets M = θ /θ′. This means that
the lateral magni cation is expressed by the ratio between the angles made with the optical axis by
the same ray in the object and image spaces (ray in red in gure). The ratio Mθ = θ’/θ is called Angu-
lar Magni cation. Therefore, from the optical invariant one gets that this parameter is the inverse of
the Lateral Magni cation:

17














fi


fi
fi

fi

fi

fi

fi
1
Mθ = .
M
From a practical point of view, assuming that the ray drawn in red in Figure 1.15 is a marginal ray,
being θ half the angular aperture of the system, and θ’ half the angular aperture on the image side,
the Lagrange invariant tells us that the products [size × angular aperture] in the object and image
spaces are equal to each other. This means that, if a lens magni es an object, the angular aperture of
the bundle of rays forming the image shrinks with respect to the one entering the lens. Vice-versa, if
the image is smaller than the object, the aperture of the rays in the image space widens. We will see
that the Lagrange invariant has a counterpart in radiometry when we will consider the relationship
between the radiance of an image and that of an object. Moreover, when dealing with optical bres,
where light is con ned both in angle (the numerical aperture) and space (the core size) the Lagrange
invariant has important implications. In fact, the light exiting the tip of a bre can be e ciently cou-
pled into a bre of similar or larger core, whereas trying to couple that light into a bre of smaller
core could result in severe losses (see section 6.2 and exercise 4).

THE ABBE’S SINE CONDITION

An optical system is said to be stigmatic for two axial points P e P’ if P’ is a perfect replica of P
(i.e. the system does not have spherical aberration, as shown in Figure 1.16. Actually, to achieve
such a result with a single lens, this has to be aspherical.
A simple spherical lens would in fact gives the image shown in Figure 1.17, where the rays con-
verging toward P’ cross the optical axis in di erent positions. This is due to the fact that the annular
regions of the lens have a di erent Dioptric Power.

Figure 1.16 Image of an axial point P made by an Figure 1.17 Aberrated image of an axial point P
aspherical lens. made by a spherical lens.

Images made by a stigmatic system might still be a ected by coma. This aberration a ects o -axis
points and is named from the shape of the blur a ecting such points, which resembles a comet tail.
Coma depends on the uneven linear magni cation produced by the di erent annular regions of a
lens. An example of image a ected by coma is shown in Figure 1.18.
In order for a stigmatic system to be free of coma (o -axis points are perfect replicas of each other)
the Abbe’s sine condition, i.e the equation:

n h si n(θ ) = n′h′si n(θ′) (1.2.13)

18


fi
fi

ff
ff

ff
fi
ff
ff
ff

fi

ff
fi
ff
fi
ffi

ff
fi
must be satis ed by all the rays leaving the optical axis. In equation (1.2.13) n and n’ are the re-
fractive indices in the object and image spaces, h and h’ are the object and image heights and θ and θ’
are the slope angles of axial rays in the object and image spaces, respectively.
The Abbe’s condition requires that the size of the object h and the image h’ are small, but the angles
θ and θ’ can be arbitrarily wide.

Figure 1.18 Coma (from the http://www.olympusmicro.com website)

A system that is stigmatic for two point P and P’ and satis es the Abbe’s condition is said to be
aplanatic. Such a system is free from the most important aberrations, but the chromatic aberration,
which is not considered here.
We will shortly see that the absence of aberrations is required to describe with a simple formalism
the transport of energy by a radiation eld trough an optical system. Therefore, in radiometry, we
assume that the Abbe’s sine condition is always met.

Figure 1.19 Aplanatic optical system.

19
fi

fi

fi

Section 3

THE RADIOMETRIC QUANTITIES


Radiometry is the branch of physics that teaches how to measure the power/energy content of a
radiation eld. More speci cally, radiometry studies the propagation of radiant power/energy in free
space and in optical systems. Radiometry mainly refers to incoherent sources and assumes that light
propagation obeys the laws of geometrical optics. This means that di raction e ects and aberrations
in optical systems are neglected.
Even if an appropriate theory of radiometry should consider the spectral properties of radiation
elds and the behaviour of optical media as a function of the wavelength, in the following we will
consider only the total power carried by the radiation, all over its spectral range. Actually, many (qua-
si) transparent media, including the atmosphere, present absorption bands that can remove photons
from the radiation at speci c wavelength, in such a way that the power and the spectral content of a
beam change along its path. Yet, coherently with the assumption of neglecting the spectral features of
radiation elds, we will assume that they propagate only in lossless media, i.e. in media that do not
alter the power content of the beam.
Radiometry is based on a set of physical quantities listed in Table 2.1. Thereafter, the most impor-
tant radiometric quantities are de ned.

Table 2.1. Radiometric quantities

Quantity Symbol De nition Unit

Radiant energy Q ∫Φ dt J

Radiant energy density u dQ/dV J/m3

Radiant ux Φ dQ/dt W

Radiant exitance M dΦ/dA W/m2

Irradiance E dΦ/dA W/m2

Radiance L dΦ2/(dApr⋅dΩ) W/(m2·sr)

Radiant intensity I dΦ/dΩ W/sr

Although for a light pulse the appropriate radiometric quantity to describe its radiation eld is the
Radiant energy, hereafter we will consider only Continuous Wave (CW) electromagnetic elds or
elds with a slowly varying average power, so that the quantity of choice to start with, is the power of
the beam.

Radiant Flux
The Radiant ux Φ is the power carried by a radiation eld. The eld is supposed to be emitted by
a source and/or received by a surface. The radiant ux is a fundamental quantity in radiometry and is
measured in watts (W).

20
fi
fi
fi
fl
fi
fi

fl

fi
fi

fi

fl
fi
fi

ff
ff
fi
fi
Irradiance
The Irradiance E(x,y) is the density of radiant ux received per unit area by a real or virtual sur-
face (Figure 1.20 a):


E(x, y) = . (1.3.1)
dA
The radiation ows inward onto the surface with no regard to incoming angular density. The Irra-
diance E(x,y) is a function of the position (x,y) on the receiving surface and is measured in W/m2.

Radiant Exitance
The Radiant exitance M(x,y) is the density of radiant ux emitted by an extended source per
unit area:


M(x, y) = . (1.3.2)
dA
The radiation ows outward from the surface with no regard to angular density (Figure 1.20 b).
In general, the radiant exitance M(x,y) of a primary source of radiation or of a body that di uses/re-
ects/transmits the radiation emitted by a primary source is a function of the position (x,y) and is
measured in W/m2.

Radiant Intensity
The Radiant Intensity I(ϑ,φ) is the density of radiant ux per unit solid angle emitted by an en-
tire source in a given direction (Figure 1.20 c). The radiation ows outward from the source with no
regard for surface area. The radiant intensity is de ned as:


I(ϑ, ϕ) = . (1.3.3)

It is a function of the direction de ned by the polar angles ϑ,φ, and is measured in W/sr. Although a
radiant intensity can be de ned for any source, this quantity is mostly used to characterise point-like
emitters, i.e. sources that are intrinsically small or whose size is much smaller than the distance to
the detector.

Radiance (or Brightness)


The Radiance L(x,y; ϑ,φ) is a very important quantity in radiometry and represents the density of
radiant ux per unit solid angle and unit projected area along a speci c direction (Figure 1.20 d).
The Radiance is a property of an electromagnetic eld propagating in a lossless medium, irrespective
of spatial location. Considering a reference surface, which can be either an extended emitter (source),
a receiver (detector) or whatsoever surface in between, the radiance is de ned by the equation:


L (x, y; ϑ, ϕ) = . (1.3.4)
d Aproj.d Ω

It is a function of the position (x,y) and of the direction de ned by the polar angles (ϑ,ϕ), and is mea-
sured in W/(m2·sr). For a generic source, the radiance may comprise light emitted by a luminous
body, re ected, transmitted, or any combination thereof.
In should be noted that in equation (1.3.4) the area to be considered is the projection of the actual
area of the reference surface along the direction of sight, described by the angles (ϑ,ϕ) with respect to
the normal. This means that Eq. (1.3.4) can be rewritten as:

21
fl
fl
fl

fl
fl



fi

fi
fl
fi
fi
fl

fl
fi

fl
fi
fi

ff


L (x, y; ϑ, ϕ) = , (1.3.5)
d A d Ω cos(ϑ)

where ϑ is the angle between the normal to the surface un and the direction of sight, as depicted in
Figure 1.20 d.
The Radiance describes the distribution of the power carried by a radiation eld, as a function of
the position and the propagation direction, with reference to any surface and its normal.

Figure 1.20 Graphic representation of radiometric quantities: (a) irradi-


ance, (b) exitance, (c) intensity, and (d) radiance. From “Cornelius. J.
Willers, Electro-Optical System Analysis and Design”

22

fi

Section 4

LAW OF CONSERVATION OF THE RADIANCE


The radiance is the most fundamental quantity of radiometry, because it is a property of the radia-
tion eld that is independent on the propagation channel, provided that no photon loss takes place
due to either absorption or scattering. In fact, an important theorem states its conservation.

Free space propagation


Let’s assume that a radiation eld propagates in a homogeneous, non absorbing medium, from a
source to a receiver. The Etendue of the radiation eld emitted by an in nitesimal portion of the
source of area dA0 and con ned to a solid angle dΩ0, forming an angle ϑ0 with the normal to the sur-
face, is de ned as:

dℰ0 = d A0 cos ϑ0 d Ω 0 (1.4.1)

Figure 1.21 Transmission of radiant ux between a source and a


receiver

The etendue is a purely geometrical quantity that represents an invariant of the beam (for compari-
son see the Lagrange invariant in subsection 2.3). The elementary ux dΦ transmitted by the in-
nitesimal source dA0 is given by the product of the etendue with the radiance L of the source [
dΦ = L dℰ], in such a way that the radiance can be expressed by the equation:

L = . (1.4.2)
dℰ
We further de ne the solid angle dΩ0 subtended by an element dA1 of the receiver surface, as seen
from the source:

d A1 cos ϑ1
d Ω0 = , (1.4.3)
r2
and the solid angle dΩ1 subtended by the element dA0 of the source surface as seen from the receiver:

23
fi
fi

fi
fi



fi
fi
fl
fi
fl
fi

d A0 cos ϑ0
d Ω1 = . (1.4.4)
r2
From Eq. (1.4.1) and (1.4.2) it is evident that the element of ux dΦ emitted by the source and
collected by the receiver is given by:

d Φ = L 0 dℰ0 = L 0 d A0 cos ϑ0 d Ω 0, (1.4.5)

where L0 is the radiance measured at the source. Using Eq. (1.4.3), we can write a general expression
for the ux transferred along the transmission channel:

L 0 d A0 cos ϑ0 d A1 cos ϑ1
dΦ = . (1.4.6)
r2
Using the de nition in Eq. (1.3.4), the radiance at the receiver is then:

dΦ L 0 d A0 cos ϑ0 d A1 cos ϑ1 r2
L1 = = ⋅ = L 0, (1.4.6)
d A1 cos ϑ1 d Ω1 r 2 d A1 cos ϑ1 d A0 cos ϑ0

which is the same as the radiance at the source.


This result might appear surprising since we experience everyday that the light emitted by a bulb
spreads with propagation. Indeed, in section 1.1 we have seen that the intensity of an electromagnet-
ic wave, de ned through the Pointing vector, decreases with the distance from the source.
Let’s consider an extended source and a small portion dAS on it. The light exiting dAS is captured
by a detector of area dAD positioned at two di erent distances from the source (Figure 1.22). As ex-
pected, less light is collected by the detector when we move away from the source, but, also the solid
angle dΩ subtended by the element of area dAS at the detector decreases. Owing that the radiance is
calculated by dividing the power at the receiver by dΩ, we should’t be surprised that the radiance of a
radiation eld does not change with propagation.

Figure 1.22 Radiance of a light beam measured at


different distances from the source.

24
fl
fi
fi
fi



ff

fl

When we watch an object, the amount of light captured by our pupils diminishes with distance,
but the solid angle under which the object is observed decreases as well, as shown in Figure 1.22.
This demonstrates that our eyes measure the radiance L(x,y;ϑ,φ) of the scene around us.
Finally, from Eq (1.4.2), we note that the constancy of the radiance requires that the radiant ux
dΦ does not su er any loss and that the etendue is a property of the light beam (invariant). In fact,
the etendue can never decrease, like the entropy of an isolated thermodynamic system. Light beams
transport energy, hence we should expect that their propagation obey the principles of thermodynam-
ics. In fact, the decrease of the etendue of a beam would result in a violation of the second principle
of thermodynamics.

Interface between two optical media


When a radiation eld falls onto the interface between two optical media with di erent refractive
indices, the wave front and the rays underneath undergo refraction, following the Snell’s law. Let’s
consider two media with refractive indices n1 and n2 and a small surface element dA at the interface Σ.
A narrow pencil of rays of radiance L1 crosses dA from the side n1, forming an angle θ1 with the nor-
mal to the surface Σ. The radiant ux carried by the beam is given by:

d Φ = L1 d A cos(θ1) d Ω1 (1.4.7)

where dΩ1 is the solid angle subtended by the beam at the surface.
The refracted beam forms the angle θ2 with the normal to the surface and subtends the solid angle
dΩ2, still carrying the same radiant ux dΦ (Figure 1.23):

d Φ = L 2 d A cos(θ2 ) d Ω 2 (1.4.8)

Figure 1.23 Refraction of a light beam at the interface between


two media.

To determine the radiance L2 in the medium of refractive index n2, we need to consider the change
undergone by the solid angle dΩ across the interface between the two media. In spherical coordinates
(Figure 1.24), the ratio between the solid angles dΩ1 and dΩ2 is given by:

d Ω1 si n(θ1) d θ1 d ϕ1
= , (1.4.9)
d Ω2 si n(θ2 ) d θ2 d ϕ2

where dϕ1 and dϕ2 are elements of the azimuthal angle, as shown in Figure 1.24.

25
ff
fi




fl
fl

ff

fl
The Snell’s law states that a refracted ray must remain in the plane of incidence, so that the az-
imuthal angle ϕ does not change (Figure 1.25):

d ϕ1 = d ϕ2, (1.4.10)

while the slope angle θ is a ected by refraction:

si n(θ1) n
= 2 (1.4.11)
si n(θ2 ) n1

Di erentiating the Snell’s law, one gets:

d θ1 cos(θ1) n
= 2 (1.4.12)
d θ2 cos(θ2 ) n1
d θ1 n cos(θ2 )
= 2 (1.4.13)
d θ2 n1 cos(θ1)

Using equations (1.4.9), (1.4.10), (1.4.11) and (1.4.13), the ratio between the solid angles dΩ1 and
dΩ2 is expressed by:

Figure 1.24 Solid angle in spherical coordinates. Figure 1.25 Refraction of a light beam at the inter-
face between two media (from Wikipedia).

d Ω1 n 2 cos(θ2 )
= 22 (1.4.14)
d Ω2 n1 cos(θ1)

Finally, using equations (1.4.7), (1.4.8) and (1.4.14), the ratio between the radiance in the two media
turns out to be:

26
ff



ff

L2 cos(θ1) d Ω1 n2
= = 22 . (1.4.15)
L1 cos(θ2 ) d Ω 2 n1

This demonstrates that the basic radiance L/n2 is preserved under refraction. Since the transmis-
sion of a radiation eld through any optical system can be modelled as a sequence of free propaga-
tions and refractions, we can conclude that the basic radiance of a beam is not a ected by optical
transformations, provided that all the media encountered by the radiation are lossless, i.e. free from
absorption and scattering.

L1 L2 Ln
= = . . . = (1.4.16)
n12 n 22 nn2

It is interesting to observe that the photon losses caused by absorption or di usion (scattering)
lower the radiance of a beam, while the presence of apertures in optical systems does not a ect the
basic radiance, even if they reduce the radiant ux.

27
fi


fl

ff
ff

ff
Section 5

LAMBERTIAN SOURCES
It is a common experience that the brightness of a sheet of paper or the appearance of a painted
wall do not depend on the observation direction or, to be more precise, on the angle formed by the
line of sight with the normal to the surface. Actually, this property is rather widespread amongst the
luminous surfaces we are likely to come across in everyday life. Nevertheless, there are surfaces that
behave in a di erent way. This is the case, for example of road signs that, mainly at night when illu-
minated by headlight projectors of cars, appear much brighter when observed within a speci c nar-
row angle. Other examples are provided by almost all metallic surfaces.
These examples refer to secondary light sources that emit light upon illumination by a primary
source, which might be the sun or any other man-made light emitter. Nevertheless, also many prima-
ry sources behave in the same way, being angle independent. This happens for uorescent lamps, for
the sun itself and for many other lighting devices, like LED based bulbs.
All light sources, either primary or secondary, that appear alike under any observation angle are
called Lambertian and share the same important property: their radiance does not depend on the an-
gles θ and ϕ:

L (x, y; θ, ϕ) = L (x, y). (1.5.1)

When the source is uniform enough to make negligible any spatial variation, the radiance is simply
a constant:

L (x, y; θ, ϕ) = L = con st. (1.5.2)

To demonstrate that the radiance L(θ,ϕ) of a lambertian source (that is a source that appears alike
from any angle) must not depend on the polar angle θ, we can use a simple argument. Let’s consider
an extended lambertian source behind a screen with a small aperture of area dA0 and a detector of
area dA1, which might be our eye, positioned along the normal to the screen (Figure 1.26).
The element of ux dΦ emitted by the source and received by the detector, according to Eq.
(1.3.4), is given by:

d A1
d Φ = L (θ, ϕ) d Aproj d Ω = L (θ, ϕ) d Aproj , (1.5.3)
r2
where dA1 is the area of the detector and r is its distance from the source. The source can be rotated
by changing the angle θ. Yet, it is clear that, being the projected area of the source equal to the area of
the aperture dA0 for whatsoever value of θ, the received element of ux is:

d A1
d Φ = L (θ, ϕ) d A0 . (1.5.4)
r2
Therefore, the requirement of a ux dΦ constant over θ calls for a radiance independent on θ.
Then, the same argument can be used to demonstrate the independence of the radiance on ϕ by rotat-
ing the source by any angle around its normal. Hence, for a lambertian source L(θ,ϕ) = L.

28

ff
fl




fl

fl

fl

fi
Figure 1.26 Independence on the angle for the radiance of a lambertian
source.

The radiant intensity emitted by a lambertian source of small area A0 or by a small portion of a
large source is given by the equation:


d Ω ∫A
I(θ, ϕ) = = L cos(θ ) d A = L A0 cos(θ ) = Io cos(θ ), (1.5.5)
0

where I0 is the radiant intensity emitted along the normal to the surface. It is worth noting that the
radiant intensity does not depend on the azimuthal angle ϕ, but only depends on the polar angle θ,
following the Lambert's cosine law:

I(θ ) = I0 cos(θ ) = L [A0 cos(θ )] (1.5.6)

Figure 1.27 Lambert's cosine law.

This speci c behaviour of the radiant intensity is the key factor to explain the property of lambert-
ian sources. In fact, Eq (1.5.6) tells us that the intensity emitted by a source in whatsoever direction
scales with the apparent area of the source A0 cos(θ), as seen from that direction. Considering the
vision process, we note that both the footprint of the source on the retina and the ux captured by
the pupil, which is proportional to the intensity emitted along the line of sight, scale with the same
factor cos(θ). This results in a constant irradiance on the retina, leading to an alike visual stimulus
for any angle (see section 8).
The radiant ux emitted by the source be can calculated by integrating the radiant intensity over
the 2π hemispace:

2π π

∫2π ∫0 ∫0
2
Φ= I0 cos(θ ) d Ω = I0 dϕ cos(θ ) si n(θ ) d θ = π I0. (1.5.7)

29
fi
fl



fl

Figure 1.28 Photometria by Johann Heinrich Lam-


bert (1728-1777).

Therefore, if the emitted ux Φ is known, the intensity I0 along the normal to the surface is:

Φ
I0 = . (1.5.8)
π
Equation (1.5.8) can be compared with the radiant intensity emitted into a hemispace by an isotropic
source, which is given by:

Φ
I0 = . (1.5.9)

A pictorial view of the Lambert's cosine law, representing the behaviour of the radiant intensity as
a function of the polar angle θ, is shown in Figure 1.29. The length of the segment from the source
to the circle (or sphere) of diameter I0 gives the the intensity I(θ).
In practice, to face most of the calculations involving lambertian sources, one should remember the
following very simple expressions of the radiant ux in terms of radiant intensity, radiance and radi-
ant exitance:

Φ = π I0 = π [L A0 ] ⇔ I0 = L A0 (1.5.10)

Φ = M0 A0 = [π L] A0 ⇔ M = πL (1.5.11)

30


fl



fl

Figure 1.29 Lambert's cosine law (from Wikipedia)

Irradiance produced by extended Lambertian sources

LAMBERTIAN DISK

Since most of the light sources we are likely to encounter in the real world are lambertian, it is
important to determine the irradiance produced by a lambertian source on a distant object. To make
things simple, let’s consider a disk of radius R and radiance L and a small detector at distance z along
the axis of the disk (Figure 1.30). We can assume that the radiant ux Φ emitted by the disk is
known. Then, the radiance L is also known:

Φ Φ
L = = 2 2 (1.5.12)
πA π R
To calculate the irradiance onto the detector we must consider the fraction of the radiant ux emit-
ted by the disk that falls within dA1. To this purpose, taking advantage of the cylindrical symmetry of
the problem, the disk can be decomposed in thin rings of area dA0:

d A0 = 2π r dr. (1.5.13)

Figure 1.30 Irradiance produced by a lambertian


disk (modi ed from “R.W. Boyd - Radiometry and
the Detection of Optical radiation”)

31
fi


fl
fl

The radial coordinate r can be conveniently expressed as a function of the distance z of the detector
from the disk and the angle θ made with the axis by the line of sight from any point in dA0 to the de-
tector:

r = z t a n(θ ). (1.5.14)

By di erentiating Eq(1.5.14) with respect to θ one gets:

z
dr = d θ. (1.5.15)
cos 2(θ )

Substituting Eqs. (1.5.14) and (1.5.15) into Eq(1.5.13), the area dA0 becomes:

t a n(θ ) si n(θ )
d A0 = 2π z 2 d θ = 2π z 2 d θ. (1.5.16)
cos (θ )
2 cos 3(θ )

From each point of the area element dA0, the detector subtends a solid angle element dΩ0:

d A1 cos(θ )
d Ω0 = , (1.5.17)
h2
where h is the distance of each point within dA0 to the detector, which can be expressed as:

h = z /cos(θ ), (1.5.18)

so that:

d A1 cos(θ ) d A1 cos 3(θ )


d Ω0 = = . (1.5.19)
(z /cos(θ )2 z2

Using Eqs (1.3.4). (1.5.16) and (1.5.19), the radiant ux transferred from dA0 to the detector dA1 is:

si n(θ ) d A1cos 3(θ )


d Φ = L d A0 cos(θ ) d Ω 0 = L 2π z 2 d θ cos(θ ) (1.5.20)
cos 3(θ ) z2

Then, the ux transferred from the whole disks to the detector can be calculated by integrating Eq.
(1.5.20) over the semi-aperture angle θ½ subtended by the radius R from dA1:

θ1/2

∫0
d Φ = 2π L d A1 cos(θ ) si n(θ ) d θ = π L d A1 si n 2(θ1/2 ) (1.5.21)

Finally, the irradiance turns out to be:


E= = π L si n 2(θ1/2 ), (1.5.22)
d A1

dΦ R2 R2
E= = πL 2 = M . (1.5.23)
d A1 R + z2 R2 + z2

Equation (1.5.22) can be simpli ed in two important cases, that is when the distance z is either
much greater or much smaller than the radius R.

32
ff

fl





fi

fl

Φ
When z ≪ R ⇒ E→M= (1.5.24)
A0

R2 π L R2 LA I Φ
When z ≫ R ⇒ E→M = = 2 0 = 02 = , (1.5.25)
z 2 z 2 z z π z2

where we have used the relationships M = π L and I0 = L A0.


In the rst case, the detector is so close to the source that the irradiance equals the radiant exi-
tance. This result can be understood on the basis of the thermodynamic equilibrium. In the second
case, the source appears so small, when seen from the detector, that it can be modelled as a quasi-
pointlike emitter, whose radiant intensity obeys the cosine law. Then, if the radiant ux ϕ or the radi-
ant intensity I0=ϕ/π are known, the irradiance onto the detector is simply obtained through the in-
verse square law approximation. As we have already seen, this law states that the irradiance generat-
ed by a pointlike (or small) source decreases with the square of the distance from the source. When
the detector is placed along the normal to the source, as it is the case in our previous example, the
appropriate radiant intensity is I0, otherwise, if the detector is placed at an angle θ to the normal, as
shown in Figure 1.31, the irradiance is given by:

dΦ I cos θ d Ω I cos θ d Ad cos θ I0 cos2 θ


E= = o = o ⋅ = , (1.5.26)
d Ad d Ad d Ad r2 r2

where r is the distance of the detector surface dAd from the source As and the factor cos2θ in Eq.
(1.5.26) accounts for the reduction of the radiant intensity with the angle θ and for the projection of
the detector area along the line of sight.
By replacing the distance r of the detector measured along the line of sight with the distance z of
its plane from the source, one alternatively gets:

dΦ I0 cos4 θ L As cos4 θ
E= = = -. (1.5.26a)
d Ad z2 z2

Finally, Figure 1.32 shows the plot of the ratio E/M from Eqs. (1.5.23) and (1.5.25) vs. the dis-
tance z, normalised to the radius R. It is evident that beyond a distance greater than three times the
radius of the source, the inverse square law can be safely applied.

Figure 1.31 Irradiance produced by a lambertian Figure 1.32 Comparison between the irradiance
source onto a detector off-axis. produced by a lambertian source onto a distant
detector and the inverse square law.

33
fi



fl

LAMBERTIAN SPHERE

We wish to determine the irradiance E on an area element dA positioned at a distance z from the
center of a spherical lambertian source of radius R and uniform radiance L all over the surface. Al-
though it is possible to determine the irradiance by performing an integration over the surface of the
source, the calculation is greatly simpli ed if explicit use is made of the spherical symmetry of the
emitted radiation. The radiant exitance of a lambertian source is given by M = π L. Thus, the total ux
emitted by the source is given by this quantity multiplied by the surface area of a sphere:

Φ = 4π R 2 M = 4π 2 R 2 L. (1.5.27)

At any distance z from the center of the sphere (z > R), this ux must uniformly distribute over a
spherical surface of area S = 4 π z2. Thus, the irradiance at distance z is given by:

π R2 L
E= . (1.5.28)
z2

This result shows that the irradiance obeys the inverse square law for any distance z. When z ≈ R
we get E → M, as it it is the case when getting close to a lambertian disk.
The radiant intensity of a sphere can be determined by observing that the emitted radiation must
be isotropic. Thus one gets:

Φ 4π 2 R 2 L
I= = = π R2 L (1.5.29)
4π 4π
Comparing Eq. (1.5.28) with Eq. (1.5.22) we note that if an observer at the position of dA (see
Figure 1.33) were to look at the sphere, this would appear as a uniform disk of radius R’ subtending
the half angle θ1/2 given by:

R
θ1/2 = si n −1 , (1.5.30)
z

where R is the radius of the sphere.

Figure 1.33 Lambertian sphere with uniform radiance.

Finally, through a simple trigonometric argument, we can note that the radius R’ of the disk di ers
from that of the sphere, and is given by:

34



fi
fl

ff
fl
R z2 − R2
R′ = R cos (θ1/2) = . (1.5.31)
z

Yet, when z ≫ R, we simply set R’ = R. This is the case for the moon or the sun observed from the
earth.

INTEGRATING SPHERE

In science and technology the need of measuring the total power carried by a radiation eld is
rather common. When the radiation eld possesses a limited spectrum of directions, either because it
is a laser beam or an incoherent beam collimated by an optical system, the task is simple and can be
performed using a large area detector. Yet, in many cases, the light comes from a di usive medium,
which acts as a secondary source upon illumination by a primary source. In this case the object can be
approximated with a lambertian emitter and the spectrum of directions of the radiation eld extends,
at least theoretically, over 2π. Also primary sources might have a wide spectrum of directions. This is
the case of a uorescent element or a common LED, whose light is far from being collimated. Then,
the measurement of the total ux is di cult because even a large detector might not be able to col-
lect all the emitted radiation. In other cases, the source can emit over the entire space around it (4π),
which makes almost impossible to collect all the light with a single detector.
To face the common problem of measuring the total ux of an incoherent beam the best option
consists in using an integrating sphere. This is a hollow sphere with a small aperture (indeed much
smaller than its diameter) that transforms the ux entering the sphere from any direction into a uni-
form irradiance all over its internal surface, through multiple re ections. In fact, integrating spheres
have their internal surface covered with non-absorbing, highly di usive materials, like BaS04 or
polymers based on Polytetra uoroethylene (Te on®). These materials have a scattering coe cient
μs much higher than the absorption coe cient μa. This means that, using the particle theory of
light, photons entering these materials are much more likely to be di used than absorbed, in such a
way that they are eventually backscattered along a direction far from the original one. In few words:
these materials are almost perfect di users, featuring a di use re ectance coe cient r close to
100%.
The operating principle of the integrating sphere can be understood with the help of Figure 1.34.
We wish to calculate the contribution to the irradiance at a generic surface element dA’ due to any
source element dA, which is assumed to be a Lambertian source of radiance L. Using the de nition of
radiance, the ux dΦ transmitted from dA to dA’ is:

d Φ = L d A cos ϑ d Ω (1.5.32)

where dΩ is the solid angle subtended by dA’ at dA, given by: d Ω = d A′ cos ϑ′ /h 2 , where h is the
line of sight between dA and dA’ and ϑ and ϑ’ are the angles formed by h with the normal to the sur-
face elements. Due to the spherical geometry of the problem, the angles ϑ and ϑ’ are equal and the
distance of dA’ from dA is: h = 2R cos ϑ. Hence, the ux from dA to dA’ is:
d A d A′
dΦ = L . (1.5.33)
4R 2
The element of ux dΦ produces an elemental irradiance dE at dA’ given by:

dΦ dA
dE = = L 2. (1.5.34)
d A′ 4R

35





fl
fl
fl



fl
fl

fi
ff
ffi
ffi
fl
fl

fl
fl
ff
fl
ff
fl
ff

ffi
ff
fi

fi
fi
ffi
Figure 1.34 Integrating sphere

We note that this result is independent of the angle ϑ. Thus, considering that the elements of surface
dA and dA’ are generic, we can conclude that the ux leaving each portion of the spherical surface is
distributed uniformly over the reminder of the sphere. You should also remember that for a lambert-
ian source L d A = dΦ/π, so that from Eq. (1.5.33) we get:

dE = , (1.5.35)
4π R 2
which con rms the previous statement, being the irradiance given by the ux divided by the area of
the sphere.
Now we assume that a certain ux Φ enters the sphere through the entrance port and hits any re-
gion of the sphere, which becomes a lambertian source. Such a ux is di used over the entire sphere,
unless a small fraction 1-r that is absorbed, where r is the di use re ectance coe cient as already
said. Hence, the ux Φ produces the irradiance:

Φr
E1 = , (1.5.36)
4π R 2
all over the sphere. Then, the light is further di used by the receiving surface a second time, a third
time, and so on. Therefore, the actual irradiance on any portion of the sphere is given by the sum of a
series of terms:

Φ Φ r
(r + r + r + . . . ) =
2 3
E= . (1.5.37)
4π R 2 4π R 1 − r
2

The di use re ectance coe cient of BaS04 in the visible range is close to 98%, while other materi-
als (e.g. Spectralon) can reach 99%. From Eq. (1.5.37) it can be observed that the e ective ux inside
the sphere is ampli ed by the factor r/(1-r), which can be much larger than one. This is not surpris-
ing, since the ux, which is delivered in stationary regime, accumulates in what is an almost non-dissi-
pative cavity. In a real device the radiant ux absorbed by the detector, typically located at 90° respect
to the entrance port, is proportional to the total ux entering the sphere at any angle. The ba e
shown in Figure 1.34 prevent light from going straight to the detector. Eq. (1.5.37) does not consid-
er the absorption by the detector and possible extra losses in the sphere. Therefore, as a matter of
fact, the measurement of the incoming radiant ux does not relies on Eq. (1.5.37), but requires the
calibration of the sphere for every compatible detector, which might be changed according to the
needs.

36

ff
fi



fl
fl

fl
fi
ffi
fl
fl

ff
fl
fl
fl

ff
fl
fl
ff
fl
ff
ffi

fl
ffl
Section 6

RADIOMETRIC PROPERTIES OF IMAGES


Imaging is the most common application of optical systems. A wealth of imaging devices are avail-
able nowadays to enhance our perception of reality, formerly only based on our eyes, which still are
the prominent interface to the world for the human kind. These include binoculars, microscopes,
photographic cameras, smartphones and video recorders, just to mention few of them. The operating
principle of electronic imaging devices is always the same: an optical system forms a replica of the
object into a plane where a pixelated detector converts the photon ux into electron bunches that
form the image. For these devices, the most important quantity to be considered in the nal step of
image recording is the irradiance E(x, y) in the detector plane, which is transformed into a digital map
upon Analog-to-Digital (A/D) conversion of the electric charge accumulated by the detector. Never-
theless, other optical systems are designed to interface with the naked eye, like binoculars or many
microscopes. In these cases, the radiance L might be more appropriate to describe the vision process.
Even if the two quantities, radiance and irradiance, are related to each other, as we have learnt in Sec-
tion 1.3, the rst is somehow more fundamental than the second, because radiance, beyond measur-
ing the spatial distribution of the optical power, also carries information about its angular spread. In
the case of imaging systems, this corresponds to consider the solid angle subtended by the entrance
pupil of the system as seen from the object. Hence, radiance takes into account the distance of a light
source, in addition to its radiant exitance, in such a way that the appearance of a distant object re-
mains the same when moving away from it, as we have already observed in the rst section of the
present chapter.

Radiance of an image
Let’s consider the image of a small element dA0 of an object, made by an aplanatic (i.e. aberration
free) optical system. The object could be a primary source or, more commonly, a secondary source
made by a surface characterised by its own radiance, which depends both on the characteristics of the
object and on those of the radiation falling on it. Often, the surface can be assumed lambertian, but
this is not a requirement for the following discussion.
Using the de nition of radiance, the elemental ux dΦ produced by the source element
d A0 = d x 0 ⋅ d y0 at the entrance pupil of the optical system (Figure 1.35), within the solid angle
element dΩ0 is:

d Φ = L 0(θ0, ϕ) d x 0 d y0 cos(θ0 ) d Ω 0, (1.6.1)

where L0 is the radiance of the source element along the direction (θ0,ϕ).
Using the expression of dΩ0 in spherical coordinates one gets:

d Φ = L 0(θ0, ϕ) d x 0 d y0 cos(θ0 ) si n(θ0 ) d θ0 d ϕ. (1.6.2)

The ux dΦ crosses the image element d A1 = d x1 ⋅ d y1 along the direction (θ1,ϕ), being the polar
angle θ1 dependent on the angular magni cation of the system, while the azimuth angle ϕ does not

37
fl
fi
fi


fi

fl

fl

fi

fi
Figure 1.35 Radiance of an image.

change for the Snell’s law. For the conservation of the ux in a lossless system, the radiance at the
image element dA1 is:


L1(θ1, ϕ) = . (1.6.3)
d A1 cos(θ1) d Ω1

The Abbe sine condition, which holds for an aplanatic system can be used to write:

n 0 d x 0 si n(θ0 ) = n1 d x1 si n(θ1). (1.6.4)

By di erentiating Eq. (1.6.4) and substituting dx → dy for the cylindrical symmetry of the system,
gives:

n 0 d y0 cos(θ0 ) d θ0 = n1 d y1 si n(θ1), d θ1 (1.6.5)

Using Eqs. (1.6.3) and (1.6.2)one gets:

d x 0 d y0 cos(θ0 ) si n(θ0 ) d θ0 d ϕ
L1(θ1, ϕ) = L 0(θ0, ϕ) (1.6.6)
d x1d y1 cos(θ1) si n(θ1) d θ1d ϕ

Finally, using the Abbe condition, (1.6.4) and its derivative (1.6.5), a simple expression for the radi-
ance of the image is obtained:

n12
L1(θ1, ϕ) = L 0(θ0, ϕ) , (1.6.7)
n 02

L1(θ1, ϕ) L (θ , ϕ)
= 0 02 . (1.6.8)
n1
2 n0

Since the same argument can be applied to any element dx, dy of the object, we can conclude that
the basic radiance L/n2 of an image is equal to that of the object. An obvious corollary of this demon-
stration is that the image of a lambertian surface is still lambertian. Further, the basic radiance of a
radiation eld can never be increased using an optical system.
In most cases, object and image are in media with the same refractive index. Then, we can simply
say that the radiance of the image is equal to that of the object.
There are important implication within the law of conservation of the radiance. The product [size
× angular aperture] in the object and image spaces are equal to each other, as we have already ob-
served in relation with the Lagrange invariant. This means that, when the magni cation ratio is M >

38
ff

fi



fl

fi

1, the Angular Aperture on the Object Side is greater than the Angular Aperture on the Image Side
and vice-versa, when M < 1. When focusing a radiation eld onto a receiver, care must be taken to
verify that its angular acceptance is wide enough to collect most of the power carried by the beam.
As an example, we can consider an optical bre with a core size s1 and a numerical aperture NA1.
Let’s suppose that the light exiting such a bre has to be coupled into a second bre with a smaller
core size s2, and numerical aperture NA2. Severe losses would occur in this case, unless NA2 > NA1.

Irradiance of the image of a lambertian source


Collecting light from a source to a detector is a common need in photonics. Very often, this is the
rst step in the process of measuring the radiant exitance of the source, which might be a primary
source, like a LED or a secondary source like a re ecting or uorescing material illuminated by a suit-
able radiation. Digital photography and any kind of image recording process relies on appropriate
optical systems and terminate with the conversion of a fraction of the photons emitted by the subject
into electrons at a pixelated detector. The possibility to tune such a fraction of light, by changing the
angular aperture of the optics through an iris, is a strict requirement not to saturate the detector and
to optimise the signal to noise ratio, in order to achieve high quality images. On the other hand, the
precise determination of the irradiance at the detector can be used to measure the radiant exitance
from the source, provided that the angular aperture of the optical system is known. While photogra-
phy does not care much about the exitance of a light source, the measurement of this quantity is very
important in spectroscopy and in other scienti c analyses. These considerations demonstrate how
important is the capability to calculate the irradiance of an image, once the radiant exitance of the
object is known. This problem has a simple solution for a lambertian source.
Let’s consider a small element dA0 at the source, a lens, which may represent a generic optical sys-
tem, and a detector element of area dA1 in the conjugated plane (Figure 1.36).
The radiant ux dΦ emitted by dA0 in an elemental solid angle dΩ describing a cone with apex on

Figure 1.36 Irradiance of an image.

dA0, aperture θ and thickness dθ (grey shade) is:

d Φ = L 0 [d A0 cos(θ )] d Ω = 2π L 0 d A0 sin(θ ) cos(θ ) d θ, (1.6.9)

where L0 is the radiance at the source and θ is the aperture of the solid angle element
d Ω = 2π sin(θ ) d θ.
The ux collected by an optical system with angular aperture θ0 can be obtained by integrating Eq.
(1.6.9) over θ:

39
fi
fl
fl

fi
fi

fi
fl
fi

fl

fi

θ0

∫0
1/2
d Φ = 2π L 0 d A0 sin(θ ) cos(θ ) d θ = π L 0 d A0 sin2(θ01/2 ) (1.6.10)

The irradiance in dA1, which is the image of dA0, is:

dΦ d A0
E= = π L0 sin2(θ01/2 ) (1.6.11)
d A1 d A1

Taking advantage of the Abbe’s condition, we get:

n02 d A0 sin2(θ01/2 ) = n12 d A1 sin2(θ11/2 ) (1.6.12)

and nally:

n12
E = π L 0 2 sin2(θ11/2 ), (1.6.13)
n0

which, in the usual case of equal refractive indices in the object and image spaces, becomes:

E = π L 0 sin2(θ11/2 ). (1.6.14)

Equation (1.6.11) has an immediate interpretation: the radiant ux collected by the lens from area
dA0 in the source is redistributed to area dA1 in the image. Further, it is worth noting that Eq.
(1.6.14) has been already found to express the irradiance produced by a lambertian disk on a small
target on-axis (Eq. 1.5.22). Hence, we can assume that, for the on-axis elemental portion of the im-
age, the lens behaves equivalently to a lambertian disk with radiance L0. However, Eq (1.6.14) holds
for all the points of the image plane, as far as the system can be assumed aplanatic.
To recast Eq. (1.6.14) in a signi cant form, we can consider the most common de nition of the
angular aperture of an optical system, which is the Focal Ratio, also called f-number. The f-number of
a lens is de ned as:

f
f# = , (1.6.15)
D
where f is the focal length of the system and D is the diameter of the entrance pupil. It is custom-
ary to write the f-numbers “N” preceded by the symbol “f/”, in such a way to form a mathematical
expression (f/N) representing the diameter of the entrance pupil D as the ratio of the focal length f to
the number N. As an example, an optical system (e.g. a simple lens, an objective or a spectrometer)
with f-number = 2 is referred as a “f/2” optical system. This means that the entrance pupil, which in
most cases is the rst lens, has a diameter half the focal length.
For an object located at a distance s1 from the lens much larger than the focal length, using the
lens law we get that s2 is almost equal to f. This is the typical operating condition of common imaging
systems. Moreover, if the the angle ϑ1 subtended by the lens at the image plane is ϑ1 < π /8, that
1/2 1/2
is f-number N < 1, we can write:

1 1
f# ≈ ≈ (1.6.16)
2 tan(θ11/2 ) 2 sin(θ11/2 )

The above condition holds in most photographic applications, but is far from being met in mi-
croscopy, which actually use a di erent de nition for the angular aperture of the objective lenses.

40
fi

fi
fi




ff

fi


fi

fl

fi

Using Eqs. (1.6.13) and (1.6.16), the irradiance in the image plane becomes:

1 n12 π L 0 1 n12 M
E= = , (1.6.17)
4 n 02 f# 2 4 n 02 f# 2

or simply:

M
E= , (1.6.18)
4 f# 2

when the refractive indices n1 and n0 are equal to 1.


In general, when the lateral magni cation ML of the optical system is known, with simple algebraic
calculation deriving from the de nition of ML (Eq. 1.2.4), we can generalise Eq. (1.6.18) in such a
way to remove the conditions s1 ≫ f, s2 ≈ f:
M
E= (1.6.18a)
4 (1 + | ML | ) f# 2
2

which becomes equal to Eq. (1.6.18) when ML ≪ 1 , that is when the object is far away from the
lens.
If the angular aperture of the lens is wide, such as we can hardly assume tan(θ11/2 ) ≈ sin(θ11/2 ) ,
Eq. (1.6.18) is generalised to:

M
E= . (1.6.18b)
1 + 4 (1 + | ML | ) f# 2
2

While in image recording Eq. (1.6.18) is generally acceptable, in scienti c applications Eq.
(1.6.18a) and (1.6.18b) might be more appropriate. This is the case for example when a relay lens is
used to make the replica of a small primary image (e.g. at the exit port of a spectrometer) to a sec-
ondary image plane. In fact, in this case the lateral magni cation is close to 1 and the angular aper-
ture of the lens is as large as possible to limit light loss. Relay lenses are commonly embedded in
photonics instruments, like streak cameras image intensi er or gated cameras (Section 4.1).
Whichever equation is used, it is evident that the irradiance of the image is proportional to the
radiant exitance M of the source and depends on the angular aperture of the optical system f#. By in-
creasing the f-number, the irradiance of the image decreases quadratically. This is the mechanisms by
which the iris of a photographic lens controls the exposure of a picture.
It is worth noting that, for an aplanatic system, we replaced the sine of the angle subtended by the
lens at the object plane (ϑ0) in Eq. (1.6.11) with the corresponding angle in the image space (ϑ1) in
Eq. (1.6.13). This was done taking into account the lateral magni cation, that transforms any ele-
ment of area dA0 in the object space into the conjugated element dA1 in the image space. Moving the
analysis from the object space to the image space has a great advantage when considering a typical
optical system. In fact in most imaging applications, but microscopy, the distance from the object to
the lens depends on the subject, while the distance from the lens to the image is almost constant, and
close to the focal length of the optical systems. In all these cases, the irradiance in the image plane
only depends on the light emitted from the object (exitance) and on the angular aperture of the opti-
cal system (f-number), being the second quantity measured in the image space trough a very simple
equation (1.6.15). From the physical point of view, it is easy to realise why the irradiance in the im-
age does not depend on distance of the object from the lens. In fact, moving away from the lens, less
light is collected, but this distributes over a smaller area, while the angular aperture on the image

41

fi
fi

fi
fi
fi

fi

side does not change at all, and so does the irradiance, according to Eq. (1.6.14). This has to do with
the Lagrange invariant, which hold in paraxial optics, or more in general derives the Abbe sine condi-
tion that extends the same concept when the paraxial approximation is removed.
Let’s consider now the problem of capturing light to a detector, mentioned at the beginning of this
subsection. It is worth comparing the use of a lens to collect light with a ‘proximity strategy’, in
which the detector is positioned as close as possible to the source. Assuming a disk shaped lambert-
ian source, Eq. (1.6.14) holds for both the mentioned options, but the angle θ has a di erent mean-
ing, being the angle subtended either by the lens (θ11/2 in Eq. 1.6.14) or by the source (θ1/2 in Eq.
1.5.22) at the detector, respectively. Therefore, it can be easily understood that at small distance from
the source, comparable to its radius, no lens is required, whereas, getting far from the source, when
the approximation in Eq (1.6.16) holds, a large aperture lens is required to e ciently capture light.
To be more precise: for a small detector placed along the axis of the source, one should compare
Eq (1.6.18) with Eq. (1.5.23), both derived from Eq. (1.6.14). The red line in Figure 1.37 shows the
ratio E/M as a function of the distance from the source disk, normalised to its radius, for a lensless
con guration, while the blue line is the long distance approximation. These gures are compared to
the predictions of Eq. (1.6.18) for a lens of f-number f/2 or f/1.4 (straight lines). Actually, such val-
ues correspond to large aperture lenses, and can be hardly increased. It turns out that, even for the
higher f-number lens, more light is collected by the detector when it can be put at a distance from the
source lower than 2.5 times its radius (red line in Figure 1.37). Yet, as a matter of fact, a lens is of-
ten used to collect light, either because the source is small, or when it is impractical to get close to it.
In other cases, a lens is the option of choice to be able to measure the radiant exitance of the source
through Eq. (1.6.18).
Finally, we observe that Eq. (1.6.17) tells us that the irradiance of an image is also proportional to
the square of the ratio between the refraction indices n0 and n1 of the media where source and detec-
tor are in. The most important case where the refraction index n1 is higher than n0 is our own eye. In
fact, the refraction index of the humor vitreous is nHV = 1.34, close to that of water. This gives im-
ages on the retina 1.8 times brighter then they would be in air.

Figure 1.37 Comparison of the irradiance produced by a lambertian disk


on a small detector obtained either with a lens or by close proximity.

42
fi


ffi
fi

ff

Section 7

INTRODUCTION TO PHOTOMETRY
Photometry deals with the measurement of the visual response caused by electromagnetic radia-
tion with wavelength within the visible range, which is conventionally restricted to the wavelength
interval 380-700 nm, even if the sensitivity of the human eye extends beyond that limit (see exercise
12 in Section 9).
Photometric quantities, listed in Table 2, stem from the homologous radiometric quantities
“weighted” by the spectral response of the eye of normal observers (i.e. subjects not a ected by any
visual impairment). To understand the vision mechanism based on the detection of light by the eye,
we need to learn something about its anatomy and physiology.

Table 2. Photometric quantities

Quantity Symbol De nition Unit Symbol

Luminous energy Qv talbot lm·s

Luminous density uv dQv/dV lumen·s·m-3 lm·s·m-3

Luminous ux Φv dQv/dt lumen lm

Luminous exitance Mv dΦv/dA lumen/m2 lm·m-2

Illuminance Ev dΦv/dA lux lx = lm·m-2

Luminance Lv dΦv/(dApr⋅dΩ) nit lm·m-2·sr-1 = nt

Luminous intensity† Iv dΦ/dΩ candela lm·sr-1 = cd

†candela= luminous intensity produced by a source emitting monochromatic radiation at frequency ν


= 540 ⋅1012 Hz (λ = 555 nm) with a radiant intensity of 1/683⋅W/sr.

ANATOMY AND PHYSIOLOGY OF THE HUMAN EYE

The human eye (Figure 1.38) contains only a few optical components. However, in good lighting
conditions, when the pupil is small (2 to 3 mm), it is capable of near di raction-limited performance
close to its optical axis. Each individual eye also has a very wide eld of view: about 65, 75, 60, and
95 deg in the superior, inferior, nasal, and temporal semimeridians, respectively, for a xed frontal
direction of gaze. Optical image quality, while somewhat degraded in the peripheral eld of view, is
adequate to meet the needs of the neural network which it serves, since the spatial resolution of the
neural retina falls rapidly away from the fovea, which is the region of best visual acuity. Control of
ocular aberrations is helped by aspheric optical surfaces and by the gradient of refractive index in
the lens, which progressively reduces from the lens center toward its outer layers.
The main refractive elements of the human eye are the cornea and the lens, while the amount of
light reaching the retina, which is the sensitive element, is controlled by the iris. The iris is an active
membrane that resizes the diameter of the pupil from about 8 mm to 2 mm in high light conditions.
Like any optical system, the human eye can be characterised by a Point Spread Function (PSF), which

43
fi
fl

fi

ff

fi
ff
fi
Figure 1.38 Anatomy of the human eye.

is, by de nition, the circle of confusion resulting from the image of a point source in the object space.
Within the framework of Fourier Optics, the point spread function h(x,y) is the impulse response of a
2D, space invariant, linear optical system. The width of the PSF, which has typically a gaussian-like
pro le, depends on di raction and aberrations, and measures the quality of the optical system.
Breaking the assumption of space invariance, the PSF of the eye depends on the position on the
retina and on the size of the pupil. Nevertheless, still using the formalism of Fourier Optics, retinal
images can be computed by convoluting the replica of the objects, as given by the geometrical optics,
with the PSF. Speci cally, if L(x, y) is the radiance of the object, and E(x, y) is the irradiance of the
image, then:

E(x, y) = L (x, y) * h(x, y), (1.7.1)

where * represents the 2D convolution operator.


The retina is a layer of nervous tissue that covers the inside of the back two-thirds of the eyeball,
in which stimulation by light occurs, initiating the vision process. The retina is actually an extension
of the brain, formed embryonically from neural tissue and connected to the brain cortex by the optic
nerve.
The retina is a very complex structure made by several layers and cell types, like bipolar cells, gan-
glion cells and the sensitive cells, which are of two types: rods and cones. Rods and cones reside in
the innermost layer of the retina, and are deputed to convert photons in visual stimuli, while the oth-
er cells, which are part of the nervous system, constitute a sophisticated signal processing network.
Figure 1.39, which depicts the main constituents of the retina, also shows two types of photore-
ceptors: rods and cones. They play very di erent functional roles in vision: rods subserve vision at
low light levels, which is called scotopic vision, and cones at high light levels, which is called pho-
topic vision, occurring under daylight. There are three types of cones, each with a di erent spectral
sensitivity, which is the result of having di erent photopigments in their outer segment. The “long”
(L), “middle” (M), and “short” (S) wavelength cones have spectral sensitivities peaked at wave-
lengths of approximately 570, 540, and 440 nm, respectively. The information about the spectral den-
sity of the light falling on the retina is encoded by the relative activation of L, M, and S cones. L and
M cones constitute 90 to 95 percent of the total cone population; the far scarcer S cones amount to
only 5 to 10 percent. These three cone types encode the thousands of colours that humans can see.

44
fi

fi

fi
ff

ff
ff

ff

This is possible because of the broad and largely overlapping spectral sensitivities of these photore-
ceptors. The curves cover the entire visible spectrum, with the L and M cones responsible for most of
that range. All rods have the same spectral sensitivity (and the same photopigment), peaking at
about 500 nm. Therefore, scotopic vision is basically achromatic, while photopic vision is charac-
terised by color perception. Cones have the greatest density in the fovea and in the central part of the
retina and are responsible for the maximum visual acuity, while rods, which constitute some 95 per-

Figure 1.39 Schematic diagram of the retina

Schematic diagram of the retina. Note that rods,


rod bipolar cells, and rod amacrine cells are ab-
sent in the fovea (from E. R. Kandel, et al. - Prin-
ciples of Neural Science).

cent of the retina’s more than 125 million photoreceptors, are mainly distributed in the peripheral
part of the retina and are demanded to provide the highest sensitivity in darkness.

SIMPLIFIED DESCRIPTION OF THE VISUAL PROCESS

The approximate retinal irradiance for extended objects is given by the following formula:

45

Ap
E(λ) = L (λ) t (λ), (1.7.2)
278.3
where E(λ) is the retinal irradiance spectral density (W·m−2·nm−1), L(λ) is the radiance spectral
density of the object (W·m−2·sr−1·nm−1), t(λ) is the transmittance of the ocular media and Ap is the
e ective area of the entrance pupil (mm2).
Photopic retinal illuminance, Ev(λ) (cd·nm−1), is computed by an equivalent formula, provided
that the radiance spectral density, L(λ) is replaced by the luminance spectral density Lv(λ)
(cd·m−2·nm−1), de ned by:

L v(λ) = k m V (λ) L (λ), (1.7.3)

where km is a constant required to convert physical quantities into photometric quantities and V(λ) is
the standard photopic spectral sensitivity function of the human visual system. In the In-
ternational System of Units, km is equal to 683 lm/W.
A similar equation hold for scotopic vision:

L v(λ) = k′m V′(λ) L (λ), (1.7.4)

where k′m = 1700 lm /W and the function V’(λ) is the standard scotopic spectral sensitivity func-
tion of the human visual system. The functions V(λ) and V’(λ) are shown in Figure 1.40 and Figure
1.41 (normalised).

Figure 1.40 Standard photopic and scotopic spec- Figure 1.41 Normalised standard photopic and sco-
tral sensitivity function of the human visual system topic spectral sensitivity function of the human visual
(from G. Wyszecki et al. - Color Science). system (from G. Wyszecki et al. - Color Science).

Therefore, for a very simpli ed description of the conversion of light into visual stimulus, the spec-
tral sensitivity functions V(λ) and V’(λ) for the Standard Observer have to be measured. This was
done by the Commission Internationale de l’Éclairage (CIE) through experiments made with panels
of volunteers. Several methods were used. Afterwards, the Flicker method is brie y described.
In the Flicker method, a test light having a wavelength λtest is matched by the panelists to a refer-
ence light having a wavelength λref, whose brightness (or luminance) is known. The two lights are

46



ff
fi




fi

fl

alternately presented to the observers into the visual eld. At low repetition frequency, the color
ickers, for example, between red (test) to green (reference). On increasing the frequency, when a
frequency of about 30 Hz the two colours merge into one (yellow in case of green + red). Above this
frequency, no color change is perceived. However, if the two lights di er in brightness from each oth-
er, the di erence in brightness remains as a icker, even if the colours merge into one. On further
increasing the frequency to a value higher than 50 Hz, the brightness as well as the color merges to
yield a uniform visual eld. By utilising the frequency interval in which the icker attributed to
brightness di erence remains, but that for color disappears, the brightness of the test light can be
matched with that of the reference light. Using this method or other similar approaches, the spectral
sensitivity of the standard eye can be measured all over the visible spectrum, with reference to the
wavelength λ = 555 nm. The eye sensitivity at this wavelength is set through the de nition of the
photometric unit of the SI candela, in such a way that V(555) = 683 lm/W. You might wonder why
such a odd value was chosen by metrologists. This has to do with the original de nition of candela,
which referred to a standard candle burning at a precise rate. Thereafter, the de nition of luminous
intensity was changed several times, up to the modern de nition involving a monochromatic radia-
tion, yet, the amount of light corresponding to the SI unit candela was kept equivalent to that of the
original de nition.
In 1924, the Commission Internationale de l’Éclairage (CIE) established the spectral sensitivity
function for photopic vision V(λ), based on the average values measured in 7 studies involving 251
people with normal color vision. Similarly, in 1951, CIE established the spectral sensitivity function
for scotopic vision V’(λ). Although a real observer with exactly the spectral responsively illustrated in
Figure 1.40 does not necessarily exist, virtual observers with these responsivity are known as the
CIE Standard Photometric Observers.
In general, the luminous ux of a radiation eld, measured in lumen, can be calculated as:

∫λ
Φv = k m Φ(λ) V (λ) d λ, (1.7.5)

where Φ(λ) [W/nm] is the spectral density of the radiant ux. Similar equations can be written to
convert all the radiometric quantities into the corresponding photometric quantities.

THE NATURAL LIGHT AND THE HUMAN EYE

Looking at the spectral features of the human eye depicted in Figure 1.41 we might ask where
they come from. The answer is indeed very simple and can be found by considering the spectrum of a
blackbody at temperature of 5600 K, which is close to the temperature estimated for the solar corona.
The peak of the blackbody spectrum shown in Figure 1.42 closely resemble the photopic spectral
sensitivity of the human eye. Hundreds of millions of years of evolution for living organisms have
tuned the sensitivity of the eye to the light we live in.
Finally, it is worth mentioning the concept of color temperature of a light. The many points in
Figure 1.43 show the chromaticy coordinates (x, y) of the daylight, measured around the earth globe
at di erent times of the day, and in di erent seasons. The chromaticy coordinates (x, y) measure the
color of a light in a standard reference system called XYZ, where the sum of the three coordinates
x,y,z is 1. It can be noted that most of the points representing the daylight crowd around the point of
coordinates x = 1/3, y=1/3 (and z =1/3), which corresponds to the perfect white light, that is a light
with uniform spectral density all over the visible interval of wavelengths (380 nm - 720 nm). We
can observe that the points in the gure can be tted by an equation expressing the Daylight locus.
The line of the Daylight locus is close to a second line that can be plotted using the color coordinates
x, y of the light emitted by a body heated at di erent temperatures. Since the spectrum of a heated
blackbody is modelled by the famous Plank equation, this second line is called Plankian locus. There-

47
fl
ff
ff
fi
ff

fi

fl
fi

ff
fi
fl
ff
fi

fi
fi
fl
ff
fi
fl
fi

fi

fore, we learn that the color of the daylight can be well approximated by that of a body heated to a
temperature from about 3.500 K to about 40.000 K, being the temperatures that account for the ma-
jority of measurements between 5.500 K and 6.500 K. In fact, 5.500 K and 6.500 K are the conven-
tional color temperatures of the direct sunlight and the daylight that including the e ect of the blue

Figure 1.42 Comparison between the spectrum of blackbody at 5600 K


and the photopic spectral sensitivity of the human eye.

sky, respectively. Moreover, the color temperature of the perfect white light is 5.500 K. Extremely
high temperatures can be measured in the shadow, under the in uence of a deep blue sky. In general,
the color temperature is de ned also for the light emitted by arti cial bulbs. To nd the color temper-
ature of a light, one needs to nd rst its chromaticy coordinates (xT,yT). Then, the color temperature

Figure 1.43 Chromaticy coordinates (x, y) of the


natural light

Chromaticy coordinates (x, y) of the natural


light measured in different conditions on the
earth (modi ed from from G. Wyszecki et al. -
Color Science).

48
fi
fi
fi
fi
fl
fi
fi
ff
of the light can be found at the intercept between the Plankian locus and the isotemperaure line (see
Figure 1.43) passing through the point (xT,yT). The isotemperaure lines are lines orthogonal to the
Plankian locus in a uniform chromaticy diagram, that is a diagram, like the L a* b*, where the euclid-
ean distance between two points representing the colours of two lights corresponds the the percep-
tive di erence between those colours. Unfortunately, even a simpli ed theory of color is a tough top-
ic, far beyond the scope of these notes. Hence, we refer to speci c books for an extensive discussion
of color science.

49
ff

fi
fi
Section 8

CONCLUSIVE REMARKS
Radiometry is the science of measuring light, while photometry tells us how we perceive light and
colours of the world through our eyes. Radiometry and photometry are important to understand how
light propagates and is processed by optical systems, including the human visual system. But, for ex-
perimental researchers there is more than that. Photonics devices are widespread in any laboratory
operating in whatsoever research eld, from material science to biology. The Interaction between
light and matter is the most powerful tools to investigate the reality, from macroscale to microscopic
objects. Whenever light is involved, like in spectroscopy, imaging or material processing, we need to
deliver and collect photons. To properly arrange an optical setup capable of recording or measuring
light, we must rely on the principles of radiometry.
We have learned that the most important quantity of radiometry, called radiance, bene ts from a
very special property: it is invariant in lossless media, that is media where photons do not get lost
neither for absorption nor scattering. Is that condition very stringent ? Not much indeed, because the
optical glasses or the polymeric materials the lenses are made of, meet this requirement quite well.
Any conservation law is a great gift from nature. This hold for momentum and angular momentum
in mechanics or for energy in general physics. These conservation laws stem from the symmetry of
space and time and are so universal that a wealth of phenomena can be interpreted on their bases.
The law of conservation of radiance might be less fundamental, but not less useful when dealing with
radiation elds.
Never try to increase the radiance of a light beam, because you will fail. Yet, you might take advan-
tage of this conservation law to calculate what you get out of an optical system once you know what
you put in, provided that you can avoid absorption and scattering of photons.
There is an other gift in radiometry: most of the light sources we are likely to encounter bene t
from the independence of the radiance on the angle. This is part of our common experience since
most objects do not change their appearance when we turn our eyes and look at them from a di erent
direction. This property characterises lambertian surfaces and makes things simple when we need to
measure the radiant exitance of a light source, that is the power emitted per unit area. In fact, for
lambertian sources the irradiance of their images is proportional to the radiant exitance through a
factor that only depends on the angular aperture of the imaging system.
To move from radiometry to photometry we need to known the sensitivity of the human eye for
the visible part of the electromagnetic spectrum. The spectral sensitivity functions for photopic and
scotopic visions of the eye were measured several decades ago thanks to panels of volunteers free
from visual impairments. Photopic vision relies on cones, which are light sensitive cells of three
types, containing di erent absorbing pigments, in such a way to provide color perception. Scotopic
vision, instead, is based on rods, which feature higher sensitivity with respect to cones, at a price of
achromatic vision.
Further, it turns out that the sensitivity of the eye for photopic vision matches the emission spec-
trum of the sun. This nding is not surprising because there has been evolution plying around for
millions of years. Remarkably, the “color” of the daylight measured in di erent positions around the
globe and at di erent times of the day is not far from that of a blackbody at di erent temperatures.
Such a temperature is 5500 K for the direct sunlight around noon. This gives the opportunity to cate-
gorise man-made lighting devices using a unique number: the Correlated Color Temperature (CCT),

50
fi
ff

ff
fi
fi

ff

ff
fi
ff
fi

i.e. the temperature of the blackbody whose color better matches that of the light source. In fact, we
can buy light bulbs emitting cool light (CCT = 4000 K) or warm light (2700 K). Considering the
spectral sensitivity of the human eye and the emission properties of LEDs, the cool light lamps are
more e cient in terms of conversion of electrical power into luminous ux. Yet, we feel more com-
fortable with warm light, demonstrating that the principle of “economy” is not the only rule leading
the world. There is more than that: the phycologists have found that our preferences change with the
time of the day and the type of activity we are involved with. You might believe that this is not so
serious, indeed, but major manufacturers of computer systems and electronic devices (e.g. tablets and
smartphones) have recently set an automatic chance of the CCT of the screen at night. This demon-
strates once more how the quality of light is important in our life.

51
ffi

fl
Section 9

EXERCISES
1. Consider the optical system shown in Figure 1.10. For the speci c position of the object
shown in gure, the magni cation ratio is M = +1. Yet, the system has in nite couples of con-
jugated planes with different magni cation ratios.
Determine:
a) the effective focal length of the optical system;
b) the position of the object required to obtain a magni cation factor M’ = -1;
c) the eld of view of the system in the condition mentioned in question b), assuming that the
diameter of the rst lens is f and that of the second lens is 2f.

2. Find the aperture stop, the eld stop, the pupils and the windows of the optical system
shown in Figure 1.8. Comment the results.

3. A lambertian source of small size emits a radiant ux Φ. A detector of area Ad is located at a


distance d from the source along a direction that makes an angle θ with the normal to the
source. Determine the radiant ux collected by the detector as a function of Φ, d Ad and θ.

4. A researcher needs to couple the light exiting the tip of a multimode optical bre into a sec-
ond multimode bre. The rst bre has a core size W1 = 200 µm (diameter) and a numerical
aperture1 NA1  =   0.12, while the second bre has a core size W2 = 50 µm and a numerical
aperture NA2 =  0.22. Calculate the fraction of the optical power that can be coupled to the
second bre neglecting Fresnel (specular re ection) losses at the interfaces and assuming
that the optical coupler is free from aberration (aplanatic) and has a proper angular aperture.

5. A light source has the shape of a circumference of radius R, made by a very thin wire. The
source emits a radiant ux Φ isotropically. Determine the irradiance on a very small detector
located on the axis of the circumference, with its normal directed as the axis and at a distance
d = R from the center of the circumference.

1 The numerical aperture of an optical fibre is defined as the sine of half the acceptance angle: N.A. = sin (θ/2).

52
fi
fi
fi

fi
fi
fl

fi
fi
fi
fl
fi
fi

fi

fl


fl
fi

fi
fi

fi

6. A lambertian source is a circle of radius Rs = 1 cm and


emits a radiant ux Φ =1 mW. The emission is mono-
chromatic @ 500 nm and has a parabolic pro le, such
as the radiant exitance is given by the equation:
M = K (1 − r 2 /Rs2), where K is a constant and r is the
radial coordinate. To measure the emission, a re-
searcher can use a lens with aperture f/2 and a photo-
diode having a radius Rd = 1 mm. Determine:
a) the constant K;
b) the maximum radiance of the source;
c) the radiant ux that can be collected by the pho-
todiode when the image of the source has the
same radius of the photodiode.

7. In a clear night of full moon, a young Physicist, feeling romantic, wants to take a picture of his
girlfriend under the moonlight, using a CCD camera with a f/4 lens. He doesn’t trust the au-
tomatic system of the camera. Hence, he wants to personally calculate the exposure time, but
he is no good.
Please, help that poor guy to calculate the correct exposure time for his camera, knowing
that:

a) the solar irradiance on the earth/moon (also called solar constant) is 1360 W/m2, 40% of
which falls within the visible band (assume green light @ 500 nm);
b) the total diffuse re ectance (also called albedo) of the lunar surface is about 10%, while
the albedo of the human skin is about 15 % (caucasian population);
c) the distance between the earth and the moon is 3.8×108 m;
d) the radius of the moon is about 3.5×106 m;
e) the correct exposure required by the CCD to collect 104 photons per pixel, whose size is
6 x 6 μm.

8. Find the radiance of the sun from the solar constant, mentioned in the previous exercise, and
the angular extension of the sun, which is δ = 0.53 degrees when the sun is observed from
the earth.

9. Demonstrate that the solar sphere observed form the earth can be approximated as a Lam-
bertian disk.
[Hint: consider that the distance sun-earth is much greater than the radius of the sun].

53

fl


fl
fl

fi

10. A circular disk is located at a distance h from an in nitely large wall, with uniform radiance L.
The normal vector from the disk is orthogonal to the wall. Derive the equation that describe
the radiant ux owing from the wall to the disk.

11. The human eye perceives the luminance L of light sources that, for a monochromatic radia-
tion is given by Lv = 683 V(λ) L(λ) [cd/m2], where L(λ) is the radiance of the source and V(λ) is
the standard scotopic spectral sensitivity function of the human visual system, whose values
are listed in Table 1. A monochromatic source of area 1 m2 emits a radiant ux of 1 mW @
650 nm and can be assumed lambertian.
a) Considering that the minimum luminance perceived by a standard observer is Lvmin = 3
10-6 cd/m2, tell whether the light source can be seen.
b) For comparison, assume that the source is imaged by a camera featuring a CCD with
2 Mpixels (HDTV 16/9 format) in an active area A = 9.6x5.4 mm and quantum ef ciency
η = 0.6. The camera has a f/4 lens and captures 50 frames/s. The noise due to the CCD
readout process in each frame is 10 electrons/pixel. Calculate the signal in terms of elec-
trons/pixel and the Signal to Noise Ratio.
c) Comment the result.

Table 1. Spectral sensitivity of the human eye

λ(nm) V(λ) λ(nm) V(λ)

400 0.0004 560 0.995

410 0.0012 570 0.952

420 0.004 580 0.87

430 0.0116 590 0.757

440 0.023 600 0.631

450 0.038 610 0.503

460 0.06 620 0.381

470 0.091 630 0.265

480 0.139 640 0.175

490 0.208 650 0.107

500 0.323 660 0.061

510 0.503 670 0.032

520 0.71 680 0.017

530 0.862 690 0.0082

540 0.954 700 0.0041

550 0.995 710 0.0021

555 1 720 0.00105

12. Notwithstanding common belief, human eye sensitivity extends even beyond 1 µm. Table 2
gives the sensitivity of the fovea (area of maximum visual acuity) Vλ as a function of the wave-
length, normalised to the maximum eye sensitivity (Vλ = 1 @555 nm).

54
fl
fl

fi


fl
fi
A Nd:YAG laser beam, carrying the power of 1 W and having the diameter of 2 mm (assume a
at top pro le), impinges onto a perfect lambertian surface with a diffuse re ectance
R = 50%.
a) Calculate the luminance of the light source produced by the light scattered by the sur-
face;
b) assuming that the minimum perceivable luminance is about L = 3 10-6 cd/m2, determine
whether the laser light can be seen by a normal observer standing at a distance of 5 m
from the source.
c) tell whether the visibility of the source depends on its distance from the observer and
whether getting closer to the source improves its visibility.
[1 candela = luminous intensity produced by a source emitting monochromatic radiation at
λ = 555 nm with a radiant intensity of 1/683 W/sr].

Table 2. Infrared spectral sensitivity of the human eye.

λ (nm) Vλ λ (nm) Vλ λ (nm) Vλ λ (nm) Vλ

710 2.1 x 10-3 810 2.1 x 10-6 910 9.0 x 10-9 1010 7.8 x 10-11

720 1.1 x 10-3 820 1.1 x 10-6 920 5.4 x 10-9 1020 6.1 x 10-11

730 5.2 x 10-4 830 6.3 x 10-7 930 3.2 x 10-9 1030 4.9 x 10-11

740 2.5 x 10-4 840 3.4 x 10-7 940 1.6 x 10-9 1040 3.5 x 10-11

750 1.2 x 10-4 850 2.0 x 10-7 950 6.6 x 10-10 1050 2.5 x 10-11

760 6.0 x 10-5 860 1.2 x 10-7 960 3.5 x 10-10 1060 1.7 x 10-11

770 3.0 x 10-5 870 6.8 x 10-8 970 2.4 x 10-10 1070 1.2 x 10-11

780 1.5 x 10-5 880 4.0 x 10-8 980 1.6 x 10-10 1080 7.3 x 10-12

790 8.0 x 10-6 890 2.5 x 10-8 990 1.2x 10-10 1090 5.0 x 10-12

800 4.0 x 10-6 900 1.4 x 10-8 1000 1.0 x 10-10 1100 3.2 x 10-12

55
fl

fi

fl
Chapter 2

NOISE IN PHOTODETECTORS
“Since, ordinarily, channels have a certain amount of noise, and therefore a nite
capacity, exact transmission is impossible”
- Claude Elwood Shannon

1Gianluca Valentini is grateful to the colleagues of the Department of Electronic and Information of
Politecnico di Milano for giving him some notes on electronic noise from which part of this discus-
sion has been inspired.

56

fi

Section 1

INTRODUCTION
Electronic noise is a critical issue in circuit design and in signal processing. Noise is also dramati-
cally important in photodetection, since the ultimate sensitivity of any detector is set by the noise
gure through the concept of Noise Equivalent Power (NEP).
While di erent causes of noise can be considered in a extensive theory, we will mainly focus on
two sources of noise, i.e. Thermal noise and Shot noise, both of them somehow related to the quanti-
sation of the electric charge. In a more fundamental framework, noise in photodetection also stems
from the quantum nature of the interaction between radiation and matter, which underlies the con-
cept of photon.
The description of noise in the time domain can be done by considering its statistical properties.
To make things simple, we can start from the noise voltage that could be measured across the termi-

Figure 2.1 Typical waveform of the noise voltage mea-


sured at the terminals of a resistor (VRMS = 1).

nals of a resistor. Actually a Load Resistor (RL) must always be present in a detection circuit to con-
vert into a voltage signal the current generated by electrons made available by the absorption of pho-
tons.
The random motion of the free charges present in the resistor (in most cases only electrons) pro-
duces a voltage di erence across its terminals, which has the typical behaviours depicted in Figure
2.1.
The analysis of the statistical properties of noise can be done by sampling the waveform with a
digital oscilloscope. This operation results in the acquisition of many (thousands or more) time sam-
ples xi, of a random variable that, for generality, will be called X.
The mean value ⟨x⟩ and the standard deviation σx of the samples xi can be calculated with the for-
mulas:

57
fi

ff

ff

1 N
N∑
⟨x⟩ = xi, (2.1.1)
i=1

N
∑i=1 (
xi − ⟨x⟩)
2

σx = . (2.1.2)
N

Notwithstanding the random nature of noise, a proper model can be often assumed to describe its
statistical properties in a stationary system. If this is the case, the random uctuations of the quantity
under investigation (e.g. voltage, current or electron ux) can be completely characterised by a suit-
able Probability Distribution.
In a stationary process characterised by an average number of events λ occurring in a xed in-
terval of time T, the chance for the occurrence of a given number n of events in the time interval T
is described by the Poisson Probability Distribution, provided that the events are independent from
one another. This holds for many physical processes, like radioactive decay, electron tunnelling, pho-
ton emission or absorption, etc. In several cases, λ can be predicted on the basis of the success rate S
of a process and the expected number of trial Q taking place in the time interval T: λ = S/Q. In turn,
the success rate S can be predicted by a physical law or derived from the experience.
The Poisson distribution p(n) is a discrete function of the integer variable n and the real para-
meter λ, whose behaviour is depicted in Figure 2.2 for di erent values of λ:

λ n exp(−λ)
p(n) = , (2.1.3)
n!


 p(n) = 1. (2.1.4)
n=0

The probability distribution p(n) is normalised and can be used to predict the average number of
events ⟨n⟩ and the standard deviation σn resulting from a very large number of measurements carried
out on a process that obeys the Poisson statistics:


⟨n⟩ = n p(n) = λ, (2.1.5)
n=0


(n − λ)2 p(n) =

⟨σn⟩ = λ. (2.1.6)
n=0

This hold, for example, when we repeatedly count the number of photons revealed in a certain
time interval by a quantum detector that receives light from a very dim source.
While the result of Eq. (2.1.5) is somehow obvious, since λ is the expected number of events in T,
equation (2.1.6) expresses a peculiar property of the Poisson distribution, which has important con-
sequences. First of all, it is worth noting that the Poisson distribution depends on a single parameter
(λ), which is the average number of observed events ⟨n⟩ in the reference interval T. Moreover, the
Coe cient of Variation (CV) (also called Relative Standard Deviation), which is given by the ratio
between the standard deviation and the mean value and describes the instability of a non determinis-
tic process, diminishes with the square root of the expected number of events λ:

58
ffi





fl
ff
fl

fi
Figure 2.2 Poisson Probability Distribution for different values of
the expected success rate λ.

σn λ 1
CV = = = . (2.1.7)
⟨n⟩ λ λ

The Poisson probability distribution can be used to model the ux of photons carried by a light
beam or the electrons owing in a current, being both photons and electrons discrete entities. Yet, for
a process resulting from the superposition of a large number of independent elementary events
obeying the Poisson distribution, the Central Limit Theorem leads to the Normal Probability Dis-
tribution to describe the random uctuation of the physical quantity that characterises the entire
process and can be assumed to be continuos. This is the case for the power of a light beam or the cur-
rent intensity in a circuit, with reference to the above mentioned examples. The Normal probability
distribution is a continuous function of the real variable X that describes the probability density
that the random quantity X assumes a value around a given value x. For such a reason, the Normal
distribution is expressed by a Probability Density Function (PDF). Hence, PDFNor mal (x) ⋅ d x is
the probability that the quantity X assumes any value between x and x + dx in a single measurement.
The Normal PDF is expressed by a Gaussian function normalised in area and peaking at the aver-
age value of the variable X (Figure 2.3):

[ 2 ]
1 (x − ⟨x⟩)
exp −
σx2
PDFNor mal (x) = (2.1.8)
2π σx
+∞

∫−∞
PDFNor mal (x)d x = 1. (2.1.9)

Figure 2.3 depicts the histogram of the samples taken from the waveform shown in Figure 2.1,
along with the Normal PDF, having null mean value and unitary σ, which well represents the distribu-
tion of the values xi.
Once the PDF of a random variable X is known, the average value ⟨x⟩ and the standard deviation σx
present in equation (2.1.8) can be obtained through the equations:

∫−∞
⟨x⟩ = x PDFNor mal (x) d x, (2.1.10)

59

fl



fl

fl

∫−∞ (x − ⟨x⟩) PDFNor mal (x) d x .


2
σx = (2.1.11)

Figure 2.3 Histogram of the samples of Figure 2.1.


The PDF is also shown.

In many cases, it turns out that ⟨x⟩ = 0, like for the waveform in Figure 2.1 Then, the Normal
distribution is completely determined by its standard deviation σx, which gives the width of the dis-
tribution. It is easy to demonstrate that the area subtended by the PDF within the interval −σx , + σx
accounts for 68% of the samples xi resulting from repeated measurements of the variable X.
Actually, when considering noise, the assumption that the variable X, which is used to describe it,
has a null mean value, i.e. ⟨x⟩ = 0, is generally acceptable, either because the noise has intrinsically
zero average or because any CW component in a stationary process, being a deterministic quantity,
can be easily subtracted, using a proper electronic circuitry or through digital signal processing.
In practice, the mean value of the noise vanishes and the variable X, which can be either a voltage
or a current intensity, can be described by a Normal probability distribution with null average. Hence,
the only important quantity needed to describe noise is the Root Mean Square (RMS) value of the
variable X or, alternatively, its mean square value ⟨x 2⟩ , which is proportional to its average power.
The two quantities can be calculated through the equations:
+∞

∫−∞
2
⟨x ⟩ = x 2 PDFNor mal (x) d x = σx2, (2.1.12)

+∞

∫−∞
R MSx = ⟨x 2⟩ = x 2 PDFNor mal (x) d x = σx. (2.1.13)

= 0, the mean square value ⟨x 2⟩ of the variable X, which can be


It is worth noting that, being ⟨x⟩
measured from the samples xi taken from the waveform, is equal to the variance σx2 of the PDF.

60



The characterisation of noise in time domain, which leads to a function of time usually named n(t),

⟨n (t)⟩ . Yet, this


2
is extremely simple and intuitive, since it results in a single parameter: σn =
approach does not give any insight into possible strategies for its reduction. To this purpose, it is bet-
ter to move to the frequency domain. This can be done by analysing the spectral properties of the
noise, using an instrument called Spectrum Analyser. The spectrum analyser is conceptually based on
a sharp tuneable lter that is swept over the range of frequencies contained into the waveform, while
the instrument measures the mean square value of the signal passing through the lter. Such a
process, referred to a noise waveform, is schematically shown in Figure 2.4.

Figure 2.4 Operating principle of a spectrum analyser

The output of the spectral analysis depicted in Figure 2.4 is proportional to the noise power at
frequency f. Yet, its value also depends on the width Δf of the spectral lter. Dividing ⟨n 2(t)⟩|f by Δf
such a dependence is removed and one gets the noise Power Spectral Density or, more simply, the
noise Power Spectrum at frequency f:

⟨n 2(t)⟩
S( f ) =
f→f+Δ f
. (2.1.14)
Δf

Since the sinusoids are orthogonal functions, the mean square value of the sum of sinusoids is giv-
en by the sum of their mean square values. This can be proved considering two sinusoidal signals:

A2 B2
⟨[A si n(2π f1 t) + B si n(2π f2 t)] ⟩ = A 2⟨sin2(2π f1t)⟩ + B 2⟨si n 2(2π f2 t)⟩ =
2
+ .
2 2
(2.1.15)

Hence, the integral of the noise power spectral density S(f) over all the harmonic components gives
the mean square value of the noise in time domain ⟨n 2(t)⟩ , which is its average power Pn and also
the variance σn2 of the (Normal) PDF describing its statistical properties:

+∞

∫0
S( f ) d f = ⟨n 2(t)⟩ = Pn = σn2. (2.1.16)

61

fi


fi


fi

This results stems from the Parseval Theorem, which demonstrates that the integral of the square
modulus of a function x(t) is equal to the integral of the square modulus of its Fourier transform
F[x(t),ω], from which the power spectrum S(f) is obtained2.
A pictorial example of noise power spectrum is shown in Figure 2.5. The area subtended by the
curve corresponds to the mean square value of the noise ⟨n 2(t)⟩ , i.e. to its average power, which in
the example of Figure 2.5 is unitary, for compliance with the waveform depicted in Figure 2.1.
Many bene ts stem from the knowledge of the noise power spectrum S(f). First of all, its contribu-
tion to the measured signal can be reduced by ltering strategies that can take advantage of the limit-

Figure 2.5 Pictorial view of a typical noise power spec-


trum

ed bandwidth of the signal at variance with the usually much wider band of the noise. By using a
pass-band lter, properly designed to match the signal band, the noise can be attenuated by preserv-
ing the integrity of the information.

2 You might wonder why the factors 1/2 that appears in the right hand side of Eq. (2.1.15) is missing in Eq.
(2.1.16). Actually, the power spectrum S(f), by definition, is calculated considering only the positive frequencies.
This accounts for the multiplication of the square modulus of the Fourier transform by a factor 2 when the power
spectrum is calculated.

62
fi
fi

fi

Section 2

THERMAL NOISE OR JOHNSON NOISE


One of the main sources of noise in electronics and in photodetection is given by the random uc-
tuations of the free charges taking place under thermal random walk within a resistor. Such a noise is
called Thermal noise or Johnson noise and is always unavoidable in photodetection since a load resis-
tor RL has to be included in the detection circuitry in order to convert the current generated by the
incoming photon ux into a voltage signal. Further, even if no speci c resistor is intentionally insert-
ed in the circuit, the intrinsic resistance of conductive paths within the detector itself makes a load
resistor unavoidable in every radiation detection apparatus. A proper choice of the load resistor is
very important to optimise the detector responsivity, sensitivity (NEP) and bandwidth.
To model the noise generated by a resistor, we can assume that its behaviour can be described by
an equivalent circuit made of two separate components: an ideal resistor and a noise voltage genera-
tor, with a random electromotive force e(t) (Figure 2.6). The mean square value of the harmonic
noise components in the narrow frequency interval f → f+df is given by the power spectrum of e(t) at
frequency f, multiplied by the frequency interval df:

Figure 2.6 Equivalent network of a noisy resistor,


coupled to a linear network

⟨e 2(t)⟩ = S ( f )d f . (2.2.1)
f

To evaluate the noise power spectrum S(f) of the resistor at all the frequencies, we need to nd
how the noise e(t) at each frequency f is coupled to the output of a generic linear network described
by its transfer function T(jω). More speci cally, we are interested into the mean square value of the
noise ⟨e 2(t)⟩ at each frequency f that is transferred to the output of the linear network. This quantity
can be calculated with the equation:

2
⟨Ou t (t)⟩ = ⟨In 2(t)⟩
2
⋅ T ( jω) ω = 2π f, (2.2.2)
f f

which describes the In➝Out relationship that holds for mean square quantities.

63
fl

fi


fi

fl
fi
You might observe that the components of the power spectrum transfer to the output through the
multiplication by the square modulus of T(jω). The simplest, and at the same time the most signi -
cant network, is made by a single capacitor, which might represent the capacitance of the detector
1
(Figure 2.7). In this case, the transfer function of the system is T ( jω) = , and the power
1 + jω R C
spectrum at the output Sout(f), according to Eqs. (2.2.1) and (2.2.2) is given by:

⟨Ou t (t)⟩ ⟨In (t)⟩


2 2
f f 2
Sout ( f ) = = ⋅ T ( jω) (2.2.2a)
df df
1
Sout ( f ) = Sin( f ) ⋅ (2.2.3)
1 + ω2 R2 C 2
where Sin(f) is the power spectrum of the (voltage) noise generator at the input of the linear network.
Since no information is a priori available on the power spectrum Sin(f), we make the simplest assump-
tion, i.e. that the power spectrum is uniform over all the frequencies:

Sin( f ) = W, (2.2.4)

where W is a constant expressing the spectral power density of the noise.


This leads to the following expression for the mean square value of the noise voltage ⟨vJ2(t)⟩:

+∞ +∞
1 dω 1
⟨vJ (t)⟩ ∫0 ∫0
2
= Sout ( f ) d f = W =W (2.2.5)
1 + ω R C 2π
2 2 2 4 RC

⟨vJ (t)⟩ = σn = W Δ f,
2 2
(2.2.6)

1
where in Eq. (2.2.6) we have set = Δ f. Therefore, the mean square value of the noise ⟨vJ2(t)⟩,
4 RC
i.e. its average power, is given by the noise spectral power density W multiplied by the equivalent
1
bandwidth of the single pole network Δ f = .
4 RC
It is worth noting that the equivalent bandwidth Δf for the noise, being the noise expressed by a
1
quadratic function, di ers from the bandwidth for the signal, which is Δ f = .
2 π RC

Figure 2.7 Model to calculate the noise spectral


power density of a resistor

64

ff


f
To estimate the value of the spectral power density W, we can use a thermodynamic argument,
applied to the circuit shown in Figure 2.7.
The system is assumed to be in thermal equilibrium with the environment. Since it has only one
degree of freedom, having a single energy storage element, from the law of equipartition of the ener-
gy, we can assign to the capacitor an average energy equal to:

1
⟨U ⟩ = K T. (2.2.7)
2
1
Hence, being the energy stored in a capacitor given by U = C V 2, we get:
2
1 1
⟨U ⟩ = C ⟨vJ2(t)⟩ = K T. (2.2.8)
2 2
The mean square value of the noise in then given by:

KT
⟨vJ (t)⟩ =
2
. (2.2.9)
C
This means that, opening at random the switch in Figure 2.7, the charge in the capacitor (as-
sumed to be ideal) is not zero but has a mean square value equal to:

⟨Q ⟩ = C ⟨vJ ⟩ = K T C.
2 2 2
(2.2.10)

From equations (2.2.5) and (2.2.9) we nally obtain:

W = 4 k T R, (2.2.11)

which gives the appropriate value for the noise spectral power density:

S( f ) = 4 k T R. (2.2.12)

It is worth noting that the justi cation underneath equation (2.2.8), which is based on statistical
physics, does not encompass any time dependence, nor spectral feature. Hence, the assumption of
uniform power spectrum done in equation (2.2.4) is compatible with the thermodynamic approach.
Actually, even if a perfectly at frequency spectrum cannot be physically acceptable, it turns out that
the spectrum of the Thermal noise or Johnson noise is approximately at over the frequencies of
practical interest. This condition is commonly named “white noise”.
The root mean square value of the Johnson noise (white noise) passing through the bandwidth Δf
of any circuit, from Eq. (2.2.6), is therefore:

⟨vJ (t)⟩ =
2
4K T R Δ f , (2.2.13)

in term of voltage or:

4K T Δ f
⟨iJ (t)⟩ =
2
, (2.3.14)
R

in terms of current.

65








fl

fi
fi

fl

Section 3

SHOT NOISE
Shot noise is caused by the quantisation of the charge and cannot be avoided due to its fundamen-
tal origin. Shot noise arises from the random occurrence of discrete events. This is the case, for ex-
ample, when charges cross the depletion layer in a semiconductor junction.
To model the shot noise we can consider a current passing though a reverse biased pn junction,
which might be the active zone of a photodiode. If the photodiode is protected from light the current
would be only the reverse current ascribable to minority carriers (also called dark current), otherwise
the current would be mainly due to couples e-h+ generated by the absorption of photons, as we will
see in the next chapter. The pn junction is considered equivalent to a capacitor because an electric
eld is present in the region of charge depletion. For simplicity, we assume that the current is given
by a stationary ow of numerable charges (in our example electrons) through the junction (Figure
2.8).

⟨is (t)⟩ ,
2
The aim of the present discussion is to nd the mean square value of the Shot noise
which is entirely attributed to charge quantisation.
The drift of a negative charge -q within the junction induces a positive charge Q(t) = q/L x(t) on
the cathode side and a positive charge q-Q(t) = q·[L-x(t)]/L on the anode side of the junction, where
x(t) is the distance of the electron from the anode and L is the width of the depletion layer. For the
Shockley-Ramo theorem3, this results in a current i(t) owing from the cathode to the anode equal
to:

dQ(t) q
i(t) = = v(t), (2.3.1)
dt L
where v(t) is the speed of the electron. Actually, since the moving charge is negative, its passage from
the anode to the cathode corresponds to the transfer of a positive charge the other way around. In
fact, when the electron reaches the cathode, the induced charge vanishes (annihilation) and a positive
charge has been completely released from the anode to the external circuit. We assume that the elec-
tron moves at constant velocity V = L/T within the junction, where T is the transit time. Then, the
speed v(t) is given by:

L
v(t) = [θ (t) − θ (t − T )] , (2.3.1a)
T
where θ(t) represents the Heaviside function. Therefore, the electron transit induces a rectangular
current pulse i(t), from the cathode to the anode, of amplitude q/T (Figure 2.9 (a):

3The Shockley–Ramo theorem states that the current i(t) on an electrode induced by a moving point charge q is
given by:

i(t ) = q v ⋅ E(x),
where v is the instantaneous velocity of charge q and E(x) is the electric field that would exist at q instantaneous
position x under the following circumstances: the selected electrode is at unit potential, all other electrodes are
at zero potential and all charges have been removed.

66
fi

fl


fi

fl

Figure 2.8 Model for shot noise based on a pn


junction.

dQ(t) q
i(t) = = [θ (t) − θ (t − T )]. (2.3.2)
dt T
Actually, this model is rather approximated. Yet the motion of a free charge in the depletion region
of a junction can be roughly approximated as a drift with constant speed equal to the saturation ve-
locity. The saturation velocity depends on the collision of the charge with the ions in the lattice and,
under high electric eld of several kV/cm, can reach the maximum value of VeSat ≈ 107 m/s.
The overall current at the terminals of the diode can be assumed to result from the random su-
perposition of a multitude of pulses of the type represented by equation (2.3.2).
The average current ⟨i⟩ is given by the electron charge q times the mean number of carriers ⟨n⟩
that cross the junction in the unit time. The shot noise originates from the uctuations of ⟨n⟩ , ac-

Figure 2.9 Normalised current pulses with a rectangular shape


(constant speed of the charge) (a) or given by an exponential
function (b).

cording to the Poisson statistics.


To take advantage of di erential calculus, which requires derivable functions, we can assume that
the time behaviour of the current pulse carrying the charge q is given by a normalised exponential
function, shown in Figure 2.9 (b):

1 −t/τ
h(t) = θ (t) ⋅ e , (2.3.3)
τ

67
fi

ff


fl

being:
+∞ +∞
1 −t/τ
∫−∞ ∫−∞
h(t) d t = θ (t) ⋅ e d t = 1. (2.3.4)
τ

If only one electron per unit time crossed the junction on average, the probability that a current
pulse would start in the generic time interval t → t + dt would be:

p1 = d t /1 = d t. (2.3.5)

Yet, since we have assumed an average current of ⟨n⟩ electrons per second, the probability that a
current pulse starts in the generic time interval t → t + dt is:

p = ⟨n⟩ d t. (2.3.6)

To write Eq. (2.3.6) we have implicitly assumed that dt is short enough to make negligible the
probability of having more than one pulse starting in that time interval. Using a reverse time axis,
named α, to represent the delay with respect to the observation time t (forward axis), one gets that
the contribution to the current at time t given by a pulse that starts at α is equal to q h(α), where α
is the time passed since the pulse has started. On the other hand, the probability that a pulse starts in
the interval α → α + d α, according to Eq. (2.3.6) is p = ⟨n⟩ d α.
The average current ⟨i⟩ at time t is given by the contribution of all pulses starting at whatsoever

Figure 2.10 Current pulses starting at two different instants α and β.

instant α, preceding t. Hence we get:


+∞ +∞

∫0 ∫0
⟨i(t)⟩ = q h(α) ⋅ ⟨n⟩ d α = q ⟨n⟩ h(α) d α = q ⟨n⟩. (2.3.7)

This result is rather obvious because ⟨n⟩ is the average number of charges crossing the junction
per unit time, which indeed gives the average current once multiplied by q. Yet, we have learnt that
the quantity which is important for the analysis of noise is the mean square value ⟨i 2(t)⟩ of the cur-
rent rather than its average value. To calculate ⟨i 2(t)⟩ let’s consider a couple of independent pulses
starting at generic instants α and β.
The contribution to the mean square current ⟨i 2(t)⟩ of a couple of pulses starting EXACTLY at α
and β (Figure 2.10), is given by:

68

[q h(α) + q h(β )] = q h (α) + q h (β ) + q h(α)h(β ) + q h(β )h(α).


2 2 2 2 2 2 2
(2.3.8)

Yet, α and β are generic instants of time, due to the stochastic nature of the pulses. Hence, to cal-
culate their contribution to the mean square value of the current, the terms in the right hand side of
Eq. (2.3.8) have to be weighted for their probability. For the terms [q 2 h 2(α) and q 2 h 2(β )] we need
the probability for a pulse starting in α +dα OR a pulse starting in β+dβ, which is either p = ⟨n⟩ d α
or p = ⟨n⟩ d β , while for the terms [q h(α)h(β) and q h(β)h(α) ] we need the JOINT probability
2 2

for a pulse starting in α AND a pulse starting in β, which is pjoint = ⟨n⟩ d α ⋅ ⟨n⟩ d β.
To account for any possible couple of pulses starting at whatsoever instants α and β, we must use
the integral calculus, which gives the means square value of the current ⟨i 2(t)⟩ as:

+∞ +∞ +∞

∫0 ∫0 ∫0
2
⟨i (t)⟩ = 2 2
q h (α)  ⋅  ⟨n⟩d α + q 2 h(α) ⋅ h(β)   ⋅ ⟨n⟩ d α ⋅  ⟨n⟩ d β. (2.3.9)

You might wonder why only two terms out of the four in eq. (2.3.8) have been considered in eq.
(2.3.9). Actually, the two integrals in eq. (2.3.9) include an in nity and a double in nity of terms,
respectively, weighted for their probability, which are p for the rst integral and pjoint for the second
integral. Hence, the two integrals already account for all possible couple of pulses occurring in the
junction, as required to calculate the square average.
From eq. (2.3.9), considering the independence of the integration variables α and β and the eq.
(2.3.4) and (2.3.7), one gets:
+∞ +∞ +∞

∫0 ∫0 ∫0
⟨i 2(t)⟩ = q 2 h 2(α) ⋅ ⟨n⟩ d α + q 2 ⟨n⟩2 h(α) ⋅ d α h(β ) ⋅ d β (2.3.10)

+∞

∫0
2
⟨i (t)⟩ = q 2 h 2(α) ⋅ ⟨n⟩ d α + ⟨i(t)⟩2. (2.3.11)

Finally, using the well known de nition of the variance of a random variable X:


σx2 = p(x)(x − ⟨x⟩)  d x = ⟨x 2⟩ − ⟨x⟩2,
2
(2.3.12)

applied to Eq. (2.3.11), we gets:


+∞

∫0
σi2 = ⟨i 2(t)⟩ − ⟨i(t)⟩2 = q 2⟨n⟩ h 2(α) d α. (2.3.13)

We note that Eq. (2.3.11) tells us that the mean square value of the current ⟨i 2(t)⟩ is the sum of
two terms: the square of the average current⟨i(t)⟩2, which in a stationary process is a constant, and
an integral term that accounts for the random ow of charges. In general, also the term ⟨i(t)⟩ is
made by two contributes: the signal that stems from detected photons and the dark current. Even if
the dark current is certainly a disturbance and can be considered a form of noise, we do not consider
it in the discussion of shot noise. This occurs since the average dark current can be separated very
easily from the signal, being constant (i.e. a deterministic quantity) such as the signal, when the
temperature does not change. Hence, to describe the properties of shot noise, which is intrinsic in a
current ow, whichever are its sources, we need to isolate its random contribution in term ⟨i 2(t)⟩ of

69
fl

fi



fl

fi
fi



fi

2
Eq. (2.3.11) from the contribution of the signal, i.e. we must subtract ⟨i(t)⟩ from ⟨i 2(t)⟩ . This is
done in Eq. (2.3.13), under the assumption that shot noise has null average value.
In order to move from the time domain to the spectral domain, we must remember that the inte-
gral of the noise power spectrum S(f) is given by σi2. [Eq. (2.1.17)], that is:

+∞

∫0
S( f ) ⋅ d f = σi2. (2.3.14)

Hence, from eq. (2.3.13) we get:


+∞ +∞

∫0 ∫0
S( f ) ⋅ d f = q 2⟨n⟩ h 2(α) d α (2.3.14a)

The Parseval theorem applied to the impulse function h(t) and to its Fourier transform | H( jω) |
leads to:
+∞ +∞ +∞
2 dω 2 dω
∫0 ∫−∞ ∫0
2
h (t)  d t = H( jω) =2 H( jω) . (2.3.15)
2π 2π

Finally, using Eqs. (2.3.14), (2.3.15) and (2.3.7), the integral of the power spectrum S(f) is:
+∞ +∞ +∞
2 dω 2 dω
∫0 ∫0 ∫0
S( f )  ⋅  d f =  2 q 2 ⟨n⟩ H( jω) = 2 q ⟨i⟩ H( jω) . (2.3.16)
2π 2π

Since the integrals in the left and right hand sides of Eq. (2.3.16) are calculated for the same vari-
able f = ω/2π and over the entire real (+) domain [0 → ∞], we can assume the equality of the argu-
ments:

2
S( f ) = 2 q ⟨i⟩ H( jω) . (2.3.17)

Equation (2.3.17) tells us that the noise power spectrum depends on the Fourier transform of the
function h(t). Coming back to the more realistic rectangular function of duration T, instead of the
exponential function h(t), for modelling the normalised current pulse, we get:

[ ( 2) ( 2 )]
T T
h(t) = θ t + −θ t − , (2.3.18)

where h(t) has been expressed by an even function for convenience.


The Fourier transform of h(t) is then:

sin (ωT /2)


H( jω) = , (2.3.19)
ωT /2
and its square modulus is:

( 2 )
2 ωT
H( jω) = sinc2 = sinc2(π f T ), (2.3.20)

where the function “sinc” represents the cardinal sine function: sinc(x) = sin(x)/x.

70

From Eqs. (2.3.17) and (2.3.20), the amplitude of the noise power spectrum at low frequency
(f ≈ 0) is 2q ⟨i⟩, while its bandwidth can be found with a simple argument. We can assume that the
cuto frequency fc of the transfer function H(jω) is set by the rst zero of the sinc function. Figure
2.11 shows the plot of the square modulus of H(jω) as a function of the frequency f and also includes
a plot of the normalised current pulse h(t) in the time domain. The consistency of the functions h(t)
and H(jω) with the Parseval theorem (Eq. 2.3.15) can be checked by considering the integral of the
square of h(t) and that of the square of |H(jω)| in the proper domain. It is immediate to verify that
both the integrals are equal to 1/T.
A typical value for the transit time T is few picoseconds. Assuming T ≈ 1 ps, the cuto frequency
would be fc = 1/T ≈ 1 T Hz. Since this value is much greater than the frequency content of any rea-
sonably detectable signal, we can argue that the noise power spectrum (2.3.17) is at and equal to its
peak value S(0) all over the frequencies of practical interest. Hence:

S( f ) = S(0) = 2 q ⟨i⟩. (2.3.21)

This means that the shot noise can be modelled as white noise, having a spectral power density
S(f) = 2q⟨i⟩.

Figure 2.11 Power Spectrum (S(f) of the shot noise


assuming rectangular current pulses.

The mean square value of the shot noise current ⟨is2⟩ can be obtained from Eqs. (2.1.12) and
(2.3.14), considering the actual bandwidth Δf of the detection circuit:

Δf

∫0
⟨is2⟩ = σs2 = S( f ) d f.

Hence, for a detection system having a limited bandwidth Δf, using Eq. (2.3.21), we nally get the
rout means square of the shot noise current:

Δf

∫0
⟨is2⟩ = S( f ) d f = 2 q ⟨i⟩ Δ f . (2.3.20)

71
ff

fi

fl
ff
fi


Chapter 3

PHOTODETECTORS

72


Section 1

INTRODUCTION
Light detection and measurement is an ubiquitous task in science and everyday life. Shooting a
photograph implies measuring light, not to mention the vision process by humans and other living
organisms. Making a long distance phone call or internet browsing requires the combined action of a
multitude of photonics devices: transmitter, ampli ers and detectors, just to consider the most im-
portant ones. In fact, since several decades ago, photonics supports the backbone of the worldwide
communication infrastructure. Further, in science and technology, light is a common and powerful
tool to investigate the properties of materials. Therefore, generation and detection of light must be
considered indispensable expertise for any experimental researcher.
The electromagnetic spectrum extends well beyond the visible light. Photonics concerns the inter-
action of the electromagnetic eld with matter, which eventually ends up with the transfer of quanta
of energy (photons) in and out from the eld. This holds for emission, absorption and scattering of
electromagnetic radiation in the spectral range from far infrared to X-rays and beyond. In this spectral
range the theory of light-matter interaction requires either a pure quantum approach or, more com-
monly, the so called semi-classic approach that assumes quantum properties for matter and a classical
behaviour for radiation elds. In the low frequency range of the electromagnetic spectrum, which in-
cludes microwaves and radio waves, the more appropriate formalism to deal with radiation elds re-
lies on the Maxwell equations, within the framework of pure classical physics. This is outside the
scope of photonics.

Figure 3.1 Electromagnetic spectrum (modi ed from Wiki-


pedia)

Before proceeding further, it is worth spending few words on common names and acronyms given
to the di erent regions of the electromagnetic spectrum that concern photonics.
The ultraviolet region is divided in three main bands listed in Table 3.1:

73
ff

fi
fi

fi
fi
fi

fi
Table 3.1

Name Abbreviation Wavelength (nm) Photon energy (eV)

Ultraviolet A UVA 315–400 3.10 – 3.94

Ultraviolet B UVB 280–315 3.94 – 4.43

Ultraviolet C UVC 100–280 4.43 – 12.4

For the infrared region, di erent names are rather common to indicate the ranges of the EM spec-
trum. Table 3.2 reports the most widely accepted classi cation.

Table 3.2

Name Acronym Wavelength (nm) Photon Energy (meV)

Near-infrared NIR 0.75–1.4 µm 886–1653

Short-wavelength infrared SWIR 1.4–3 µm 413–886

Mid-wavelength infrared MWIR 3–8 µm 155–413

Long-wavelength infrared LWIR 8–15 µm 83–155

Far-infrared FIR 15–1000 µm 1.2–83

The present chapter is devoted to discuss the detection of a radiation eld and the measurement of
its energetic content whenever the concept of photon plays a key role. Di erent detector technologies
will be considered to cover the wavelength range where the interaction of radiation with matter is
mediated by photons with di erent energies. This diversity provides a wealth of opportunities for
physicists and engineers to design and manufacture a variety of detectors appropriate for many ap-
plications, which include, among others, spectroscopy, range nding and remote sensing.

74
ff
ff
fi
fi

fi
ff

Section 2

BASIC CONCEPTS OF RADIATION DETECTION


Radiation detectors can be roughly divided in two wide categories: quantum detectors and thermal
detectors.
d) In thermal detectors the energy of the absorbed photons is converted into heat, which determines
an increase in temperature in any material. In turn, this can result in a measurable physical e ect,
like the change in electric resistance or the generation of an output voltage, which is proportional
to the power of the optical signal.
e) In quantum detectors the interaction of photons with the active element of the detector induces
one-to-one events, like the emission of electrons in vacuum (also called external photoemission)
or the creation of electron-hole couples in semiconductors (sometime called internal photoemis-
sion). These events are “counted” by an appropriate circuitry that outputs a signal proportional to
the photon ux impinging onto the detector. The ux is proportional to the optical power only in
case of monochromatic radiation.
Both detector families present pros and cons that makes them appropriate in di erent spectral
ranges.

Thermal Detectors
In thermal detector all the incident electromagnetic radiation within a spectral range, typically
spanning from ultraviolet to far infrared, is absorbed by a proper surface, regardless of frequency or
angle of incidence (quasi-blackbody). This important feature provides a at response over an ex-
tremely wide wavelength interval (Figure 3.2). Further, the output signal is proportional to the ab-
sorbed optical power in a straight way. This property makes the measurement of the energetic con-
tent of a radiation eld very simple and direct.
On the other hand, the sensitivity of thermal detectors (see section 2.1 for a precise de nition of
this parameter) is, generally speaking, much lower than that of quantum detectors, at least in the
spectral regions where e cient quantum devices are readily available.
Yet, in many bands of the EM spectrum, thermal detectors are the only practical option, which
makes them the detectors of choice for: i) wide band Fourier Transform Infrared (FTIR) Spectroscopy;
ii) far infrared astronomy; iii) low cost intrusion detectors; iv) security scanners, and many other ap-
plications.

Quantum Detectors
The general structure of quantum detectors relies on the conversion of a photon ux into an out-
put current, which is eventually transformed into a voltage signal. Due to the discrete nature of light-
matter interaction events, the output is proportional to the number of photons absorbed by the de-
tector. This has two major consequences:
a) The spectral response has a wavelength threshold, set by the photon energy required for the quan-
tum interaction to take place.
b) The response (see section 2.1 for a precise de nition of this parameter) of a quantum detector is
wavelength dependent, featuring a quasi-linear behaviour vs. wavelength, up to the threshold
(Figure 3.2), where an abrupt drop takes place.

75

fl
fi

ffi

fi

fl

fl
fl
ff
fi
ff

Figure 3.2 Typical output signal from Quantum and Ther-


mal detectors

The wavelength dependence of the output response makes the measurement of the power carried
by a polychromatic radiation eld non trivial, since the spectrum of the incoming radiation must be
measured altogether, unless a priori assumptions are made.

Few Words on Sensitivity


The major strength of quantum detectors is their sensitivity, which can easily reach single photons.
To compare this gure with the sensitivity of a typical thermal detector we can calculate the number
of photons required to induce a measurable e ect in a micro-bolometer (Section 3.2). This technolo-
gy is employed to manufacture thermal imaging systems capable of revealing the presence of humans
or animals thought the infrared emissions of their bodies at 37 °C. Amongst the many uses in securi-
ty and safety applications, these devices are gaining interest in the automotive industry for the pro-
duction of alert systems capable of detecting pedestrians crossing roads in complete darkness.
Let’s consider a single pixel of a micro-bolometer. A possible size of the light absorbing element is
20 × 20 μm, with a thickness of 1 μm. Assuming that the element is made of silicon, even if more
performing technologies are nowadays available, the mass of the absorber is M ≈ 10-12 kg and its
thermal capacitance is CSi = 8.6 × 10-10 J/k. The minimum detectable temperature increase, with re-
spect to the background (Tbkg = 300 K), in a micro-bolometer is named Noise Equivalent Tempera-
ture Di erence (NETD) and is typically given by NETD = 50 mK. This means that an increase in
temperature of the absorber greater than 50 mK gives an output signal that can be distinguished from
the noise due to random temperature uctuations.
The peak wavelength of the radiation emitted by the human body (T=37 °C) is close to λpeak= 10
μm, which corresponds to a photon energy Eph ≈ 2 × 10-20 J. Therefore, the number of photons re-
quired to rise the temperature of the absorbing element of a microbolometer by an amount equal to
the NETD in adiabatic conditions, which indeed is a rather optimistic assumption, is:

M CSi N E T D
Nph = ≈ 109. (3.2.1)
Eph

This gure is several order of magnitude above the single photon limit.
This simple argument demonstrates that in the VIS-NIR range, where quantum devices are cheap
and e cient, there is no point in using thermal detectors, unless for special applications, where plen-
ty of power is available, like the characterisation of laser beams. Conversely, in the Mid-Wavelength

76
fi
ffi
ff
fi

fi

fl
ff

infrared (MWIR) or in the Long-Wavelength infrared (LWIR) regions, thermal detectors dominate
because quantum detectors are either expensive or not available.

Figures of merit for radiation detectors


The performances of radiation detector can be described by a set of important parameters dis-
cussed thereupon.

RESPONSIVITY

The Responsivity, sometimes also called Radiant Sensitivity, of a photodetector is given by the
ratio between the output signal and the input optical power. This parameter typically depends on the
wavelength of the radiation eld. Since the output could be equivalently expressed in terms of cur-
rent or voltage, the Responsivity of a photodetector is alternatively given by:

Iout [A] Vout [V ]


RI (λ) = RV (λ) = . (3.2.2)
Pin [W ] Pin [W ]

The Responsivity is an important parameter to design the front-end of a light detector, i.e the circuit-
ry devoted to process the signal coming out of the sensor terminals. In fact, the Responsivity provides
information on the expected electrical output in any light condition. This knowledge is required to
design the ampli cation stage, which might include an operational ampli er, possibly in transimped-
ance con guration, and an analog-to-digital conversion unit.
Yet, it is important to highlight that the Responsivity of a detector does not correspond to its sen-
sitivity, which is expressed by a di erent gure of merit, that is the Noise Equivalent Power, dis-
cussed hereafter.

FREQUENCY RESPONSE

The Frequency Response of a light detector determines its capability to correctly recover the time
pro le of a variable optical signal. A proper visualisation of this feature can by given by the frequency
behaviour of the Responsivity Rλ(f), in a logarithmic plot, as a function of the modulation frequency f
of the optical stimulus, in the manner of representation that is typical for the transfer function of an
electric linear network (Figure 3.3). Nevertheless, for many detectors the frequency performance is
given by a single parameter called Cuto frequency, fc. This is the frequency at which the output am-
plitude of a sinusoidally modulated optical signal reduces by a factor 2 , with respect to a CW signal
of equal average power.
This simple model assumes a single pole transfer function. In this case, the impulse response func-
tion (IRF), i.e. the response to a very fast light pulse h(t), is an exponential function characterised by
a time constant τ, which might be due to the capacitance C of the detector times the load resistance
RL:

( τ)
t
h(t) = Heaviside (t) ⋅ A exp − τ = RL C, (3.2.3)

where A is the amplitude of the response and Heaviside(t) is a function equal to {0 for t< 0; 1 for
t>0}.
By taking the Laplace transform of Eq. (3.2.3) one gets the transfer function:

77

fi

fi

fi

fi
ff
ff
fi


fi


H(s) = , (3.2.4)
1+s τ
where s = α + jω is a complex variable. The Responsivity of the detector in sinusoidal regime Rλ(f),
where f is the frequency of the optical signal, can be obtained by assuming s = jω = 2πjf, and taking
the modulus of the transfer function H(s):

Aτ R(0)
Rλ( f ) = | H( f ) | = = . (3.2.5)
1+ 4π 2 f 2 τ 2 1 + 4π 2 f 2 τ 2

From equation (3.2.5) it is very easy to observe that the cuto frequency is given by:

1
fc = , (3.2.6)
2π τ

in fact, according to the de nition, the Responsivity at the cuto frequency is 1/ 2 times the Re-
sponsivity for a CW signal.

R(0)
Rλ( fc ) = . (3.2.7)
2

Figure 3.3 Normalised transfer function with unitary Figure 3.4 Response of a detector to a very short
cutoff frequency light pulse.

NOISE EQUIVALENT POWER

The ultimate sensitivity of a detector is set by its noise gure. In fact, the lowest optical power that
can be detected is the one that produces an output signal comparable to the noise, such as the Signal
to Noise Ratio (SNR or S/N) is equal to 1. Below such a limit the signal is said to be “buried in
noise” and no reliable measurements can be done, unless complex ltering procedures are applied.
Since noise is typically expressed by a Root Mean Square (r.m.s. or RMS) quantity, the proper para-
meter that de ne the sensitivity, which is called the Noise Equivalent Power (NEP), refers to a sinu-
soidally modulated signal and is de ned as:

78

fi


fi
fi


fi
ff
ff
fi

NEP = optical power in sinusoidal regime that would produce a modulated signal
in a noise free detector equal to the r.m.s. noise measured from the detector with-
out any signal.
The noise power Pn of a detector is typically proportional to its sensitive area A and passband Δf,
assuming a white noise model:

Pn ∝ A ⋅ Δ f. (3.2.8)

Therefore, the r.m.s. amplitude noise, which is proportional to the square root of its power, is given
by:

Nr ms ∝ A ⋅ Δ f. (3.2.9)

Being the NEP expressed by the input power P required to give S/N = 1, from the de nition of Re-
sponsivity (Eq.) we have:

Sr ms N
N EPλ = P = = r ms ∝ A ⋅ Δ f, (3.2.10)
( N =1)
S
Rλ Rλ

where the dependance of the NEP on the wavelength λ has been considered through the responsivity.
While the NEP is de nitely the most appropriate parameter to represent the sensitivity of a speci c
detector, Eq. (3.2.10) let us understand that, to compare families of detectors based on di erent
technologies, “size” and “speed” must also be taken into account for a fair evaluation. In fact, we
might expect that large detectors are more “noisy” than small ones. On the other hand, since the
most important noise sources are “white”, the bandwidth of the detector sets the amount of noise
that passes to the output. Hence, a slightly di erent parameter, called NEP-star (NEP*
λ
), is de ned
to make this gure of merit independent on the detector size and bandwidth, which are design para-
meters largely dependent on applications.

N EPλ
N EP*
λ
= . (3.2.11)
A ⋅ Δf

Finally, since for a detector the higher is the performance the lower is its NEP, the inverse of NEP,
called Detectivity (D) is often considered a better parameter to compare the sensitivity of di erent
detector types.

1 1
Dλ = D*
λ
= . (3.2.12)
N EPλ N EP*
λ

Figure 3.5 shows the typical Detectivity D* of detectors belonging to di erent families. It is evi-
dent that the Detectivity of thermal detectors occupies the lowest part of the plot. However, the read-
er should note that, in order to reach high sensitivity in the MID and FAR Infrared with quantum
detectors, extreme cooling is required. This necessity is often non practical, unless you are in a
physics lab. Further, you might note that the upper right region of the plot is unaccessible to any de-
tector. This is because measurements are supposed to be done on earth, at an environment tempera-
ture of 300 K, which is the conventional comfort temperature.
The Spectral Power Density of a blackbody at room temperature (300 K) is given by the function
shown in Figure 3.6. It can be easily understood that the higher is the unspeci c power density due
to the thermal background, the lower is the capability to distinguish a weak signal buried in thermal
noise. Therefore, the Detectivity of any light sensor operating in any place at “room temperature” is

79


fi



fi


ff

ff
fi
fi

ff
ff
fi
fi

Figure 3.5 Detectivity of different families of detectors (source Hamamatsu


Photonics).

limited by the amount of background noise collected by the detector. It is worth noting that cooling
the detector does not eliminate the background noise, which originates from the space the detector is
in.

Table 3.1 Detectors for infrared radiation (source: Hamamatsu Photonics)

80

The inverse of the blackbody Spectral Power Density @300 K is shown in Figure 3.7. This gure
should be compared with the dashed lines marked “ideal photovoltaic”, “ideal photoconductor” in
Figure 3.5, which set the Detectivity limit.
The following Table 3.1 lists the Detectivity for the most important detector families for the in-
frared region beyond 1 μm, where silicon photodiodes (Section 4.3) and most of the available photo-
cathode materials (Section 4.1) cannot be employed.

Figure 3.6 Power Spectral Density of the radiation


emitted by a blackbody at room temperature (300 K).

Figure 3.7 Inverse of the Power Spectral Density of


the radiation emitted by a blackbody at room tem-
perature (300 K).

81

fi
Section 3

THERMAL DETECTORS
While huge di erences exist amongst thermal detectors in terms of technology, size and power range,
from kilowatts to picowatts, they are always based on the same general scheme depicted in Figure
3.8.
The radiation is absorbed by an absorbing element, whose surface is usually treated to make it
similar, as much as possible, to a blackbody. The absorbing element is characterised by a thermal ca-
pacitance Cth, that should be the smallest possible, compatibly with constraints set by mechanical or
thermal damage. The absorbing element is connected through a thermal link, with heat conductance
Gth, to a massive heat sink, having thermal capacitance CHS ≫ Cth and kept at the equilibrium tem-
perature Teq. Light absorption deposits a power W into the detector, leads to a temperature increase
ΔT, with respect to Teq., and determines a heat ow F = Gth ⋅ ΔT toward the heat sink. The response of
the detector stems from the conversion of the temperature di erence ΔT into an electrical signal,
through di erent physical mechanisms.

Figure 3.8 General structure of a thermal detector

The energy balance for the absorbing element within a short time interval dt gives a di erential
equation that describes the dynamic behaviour of the detector:

d(ΔT ) W G
W d t − Gth ΔT d t = Cth d(ΔT ) = − th ΔT (3.3.1)
dt Cth Cth

The general solution of Eq. (3.3.1) depends on the initial conditions and the temporal shape of the
light stimulus. Nevertheless, valuable insight on the expected output signal can be obtained by con-
sidering two important cases for the optical power delivered to the detector: a constant stimulus and
a sinusoidal stimulus.
Using the method of the Laplace transform, or the theory of linear di erential equation, the solu-
tion ΔT(t) of Eq. (3.3.1) for a constant optical stimulus with power W0 that starts at t = 0 (Heaviside
function shown as a dashed blue line in Figure 3.9) is given by:

82

ff
ff

fl

ff

ff

ff
Gth [ ( Cth )]
W0 G
ΔT (t) = 1 − ex p − th t , (3.3.2)

Gth [ ( τth )]
W0 t Cth
ΔT (t) = 1 − ex p − , with τth = , (3.3.3)
Gth

The time evolution of the temperature change is described by an exponential function characterised

Figure 3.9 Temporal evolution of the temperature signal ΔT


in a thermal detector in response to a constant optical
stimulus initiating at t = 0 (Heaviside function).

by a time constant τth = Cth/Gth, which sets the time required for the output signal to stabilise to the
regime value. This entails waiting for 4-5 time constants to properly measure the input power W0,
The thermal time constant τth also sets the frequency response of the detector. In fact, assuming a
sinusoidal optical stimulus superimposed to a CW component (blue dashed line in Figure 3.10):

W(t) = W0 + Wm cos(2π f t) W0 > Wm, (3.3.4)

the temperature di erence, after a transient that lasts a few time constants τth, is given by a stable
sinusoidal function exhibiting a phase shift with respect to the incoming optical signal. In general,
the phase shift is neglected and the temperature change is given by:

W0 Wm cos(2 π f t)
ΔT (t) = + . (3.3.5)
Gth G 1 + 4π 2 f 2τ2
th th

At the cuto frequency fc = 1/(2 π τth) one gets:

W0 1 Wm cos(2 π fc t)
ΔT (t) = + (3.3.6)
Gth 2 Gth

Equation (3.3.6) tells us that at the cuto frequency the sinusoidally modulated component Wm of
the input signal is attenuated by a factor 2 , with respect to the continuous wave component W0,
which is una ected. This results in a decrease of the modulation depth m = Wm/W0 in the output
with respect to the input.

83
ff
ff
ff




ff

Figure 3.10 shows the normalised [τ = 1] temperature signal ΔT for a sinusoidal input of unit ampli-
tude. We assume that the light starts reaching the detector at time t = 0. After the transient of a few
time units, the output signal shows an oscillatory behaviour, with a reduced modulation depth and a
phase lag. When the frequency reaches the cuto fc the modulated part of ΔT is attenuated by a factor
2 and the phase lag is π/4.

Figure 3.10 Temporal evolution of the temperature signal


ΔT of a thermal detector in response to a sinusoidal optical
stimulus initiating at t = 0.

The Responsivity of the detector is proportional to the temperature increase ΔT, which, in turn,
depends on the inverse of the heat conductance Gth of the thermal link. To increase the Responsivity,
it might seem reasonable to decrease Gth. Nevertheless, this will also increase the time constant τth =
Cth/Gth, resulting in a worse temporal response. Actually, the response of any single pole system is
characterised by a constant Gain × Bandwidth product: changing the parameters, what is gained in
terms of amplitude is lost in speed.
On the other hand, the time constant τ can be improved by reducing the thermal capacitance Cth of
the absorbing element. This can be done by using an appropriate material, like bismuth, which is a
metal with a very low speci c heat, and reducing the thickness of the radiation absorbing element,
down to the limit required to sustain the maximum power the detector has been designed for.

Thermoelectric Detectors
Thermoelectric detectors are amongst the rst type of thermal detectors ever developed. Nowadays
they are typically employed to measure the power of a laser beam. Figure 3.11 shows some examples
of such detectors, called thermopiles. They are commercially available in a wide range of power, from
less than one milliwatt to kilowatts. The scheme of these detectors conforms the archetype shown in
Figure 3.8. The absorbing element can be made by a metallic surface covered by a broadband absorb-
ing layer (e.g. sputtered gold). This provides high sensitivity over a wide spectral range. Yet, a very
thin absorbing layer is not suitable to measure the energy of laser pulses, which might vaporise the
lm. In this case, 1-2 mm thick glass disks with speci c doping are used to selectively absorb the in-
coming laser pulses within a small, but non negligible volume. The sensitivity and the frequency re-
sponse of detectors based on volume absorbers might be lower than those of detectors using surface
absorbers, but their strength is de nitely much higher. The heat sink is made by a massive metallic
body, with cooling features relying either on air convection (low power) or water ow (high power).

84
fi
fi

fi

fi
ff
fi
fl

Figure 3.11 Laser thermopiles

The temperature di erence between the active element and the heat sink is measured by a series of
thermocouples (Figure 3.12) that generate a voltage signal.
A thermocouple is made by two joints between wires of di erent metals or metal alloys. When the

Figure 3.12 Thermocouple and Thermopile

joints are kept at di erent temperatures, a voltage di erence arises between the tips of the wires. The
signal is almost proportional to the temperature di erence, but rather weak. Hence, several thermo-
couples are connected in series to increase the signal and provide a precise measure of the tempera-
ture di erence. Thereby the name of thermopile.
Thermocouples exploit the Seeback e ect, which was discovered by the Baltic physicist Thomas
Johann Seebeck in 1821. When a temperature gradient ∇T is maintained in a thin metal bar and no
electric current is allowed to ow, there will be an electric eld directed opposite to the tempera-
ture gradient, called thermoelectric eld. This eld leads to a potential di erence between the high
temperature (+V) and the low temperature (-V) ends of the bar. The thermoelectric eld is given by:

E = Q ∇T Q<0  or  − ∇V = Q ∇T, (3.3.7)

where the quantity Q, called thermopower or Seebeck coe cient, is de ned as the proportionality
constant between the thermoelectric eld and the temperature gradient.
The thermopower can be estimated using the Drude theory, which applies to the gas of electrons
in metals. In the Drude theory, the electrons collide with the ions of the lattice and, whichever is the
velocity before the collision, they acquire a mean velocity that depends on the local temperature.

85
ff
ff
ff
fl
fi
fi
ff
fi


ff
ff
ff

ffi
fi

fi
ff
fi

Hence, an electron whose last collision was at the generic position x will, on the average, have a ve-
locity v(x ) . In an oversimpli ed one-dimensional model, the electrons can only move along the x-
axis, so that, at any point x, half the electrons come from the high-temperature side, and half from
the low (Figure 3.13). Assuming that in Figure 3.13 heat ows from left to right (toward increasing
x), the electrons arriving at point x from the high-temperature side will, on the average, have had
their last collision at x − vx τ , where τ is the mean time between collisions, while the opposite sign
holds for electrons arriving from the low-temperature side. In the end, the mean velocity of electrons
vQ at a point x, due to the temperature gradient, is given by:

1
vxQ = [v (x − vx τ) − vx (x + vx τ)]. (3.3.8)
2 x

Figure 3.13 Seebeck effect in a metal bar.

The variation of vx(x) in Eq. (3.3.8) from x to x ± d x, being d x = vx (x) τ, can be expressed through
its di erential form:

2[ ]
1 d v (x) d v (x)
vxQ = vx (x) − x vx (x) τ − vx (x) − x v (x) τ , (3.3.9)
dx dx x

and with minor simpli cations:

vx2(x)
dx ( 2 )
d vx (x) d
vxQ = − vx (x) τ =−τ . (3.3.10)
dx

Actually, the velocity vQ is function of the temperature T, which, in turn, depends on the position x.
By explicitly expressing this dependence we can write:

d [vx2 (T (x))] d vx2 (T ) d T


= . (3.3.11)
dx dT dx
Before continuing, we must generalise the model to three dimensions by assuming
1 2
⟨vx ⟩ = ⟨vy ⟩ = ⟨vz ⟩ =
2 2 2
v , which holds for an homogeneous and isotropic material, and substi-
3

86
ff

fi
fi


fl

tuting the derivative of T with respect to x with the gradient ∇T in Eq. (3.3.11). In this way, from Eq.
(3.3.10) we get:

τ d v2
vQ = − ∇T. (3.3.12)
6 dT
According to the Drude model of the electric conductivity, the mean electronic drift velocity in pres-
ence of an electric eld E is:

eEτ
vE = − , (3.3.13)
me

where e and me are the electron charge and mass, respectively.


Since the electric current was assumed null in our model, the velocities due to the thermal gradient
and to the thermoelectric eld must be opposite to each other to cancel out.

vQ = − vE. (3.3.14)

Hence, recalling that by de nition E = Q ∇T , from Eqs. (3.3.12), (3.3.13) and (3.3.14) we get:

τ d v2 eτ 1 d v2 e
− ∇T = Q ∇T − = Q (3.3.15)
6 dT me 6 dT me

Equation (3.3.15) can be nally recast in the form:

1 d ( 2 me v )
2 1

Q =− . (3.3.16)
3e dT
This result does not depend on the relaxation time τ. Recalling that the speci c heat of a gas of inde-
pendent particles is given by:

d Ek
cv = n , (3.3.17)
dT
where n is the number of particles per unit volume, from Eq. (3.3.16) one gets:

cv
Q =− . (3.3.18)
3 ne
Drude evaluated Q by an inappropriate application of classical statistical mechanics to the theory of
solids, setting Cv equal to:

3
cv = nk , (3.3.19)
2 B
where kB is the Boltzmann constant. With this assumption, Drude found that the thermopower
should be given by:

kB
Q =− = − 4.3 × 10−5 V /K, (3.3.20)
2e

87





fi



fi
fi

fi

fi

for all the metals. Actually, the value of observed metallic thermopowers at room temperature is of
the order of microvolts per degree, a factor of 100 smaller. For example, the thermopower of plat-
inum, which is usually considered the reference, is -5.1 μV/K @300 K and changes with temperature.
This discrepancy depends of an error in Drude's model a ecting the average velocity of electrons at
thermal equilibrium, which is underestimated by a factor 10. The incongruence of the values of the
thermopower with experiments o ers unambiguous evidence of the inadequacy of classical statistical
mechanics in describing the metallic electron gas. Quantum statistical mechanics, based on the Fer-
mi-Dirac distribution is required to obtain the correct value of the thermopower Q. In fact, using the
appropriate value of the speci c heat in Eq. (3.3.10):

2 ( ϵF )
π 2 kB T
cv = n kB, (3.3.21)

where εF is the Fermi energy of the metal, one gets a fairly reasonable estimation of the thermopower:

6 e ( ϵF )
π 2 kB kB T
Q =− . (3.3.22)

In a thermocouple, with reference to Figure 3.12, the thermoelectric voltage generated within Metal
2 is brought to the terminals by the two wires made of Metal 1, which also experience a thermoelec-
tric voltage. Considering a thermocouple made of Metals 1 and 2 with thermopowers Q1 and Q2, re-
spectively, from Eq. (3.3.7) it is easy to demonstrate that the voltage di erence across the terminals
of the thermocouple is given by:

ΔV = (Q1 − Q2 ) (T1 − T2 ), (3.3.23)

provided that the terminals are kept at the same temperature.


In a thermoelectric detectors featuring N equal thermocouples from the light absorber and the
heat sink, the output voltage is given by:

ΔV = N (Q1 − Q2) ΔT, (3.3.24)

Where ΔT is the one expressed by Eq. (3.3.2) or (3.3.5) or by any solution of Eq. (3.3.1).

Bolometers
Bolometers are thermal detectors based on a thin slab of a material whose electric resistance is
highly sensitive to temperature change. The rst bolometer was invented in 1878 by the American
astronomer Samuel Pierpont Langley. Langley used two thin ribbons of platinum foil connected so as
to form two arms of a Wheatstone bridge, which is commonly used to measure small resistance
changes. One strip was shielded from radiation and one exposed to it, so that a temperature di er-
ence could be established from the sensing branch and the reference one. Bolometers can measure
radiations over an extremely wide spectral range, from X-rays to THz. However, this detection
scheme is particularly important for measuring far infrared and THz waves, because bolometers are
amongst the most sensitive detectors in this spectral range.
Even if metals can be used as well, doped semiconductor, such as Si or Ge, are typically preferred,
taking advantage of the higher temperature coe cient of these materials with respect to metals.
Around the working point of the detector, corresponding to the resistance R0 at temperature T0, the
electric resistance R of a bolometer can be assumed a linear function of the temperature T:

88




fi

ff

fi
ffi

ff

ff

ff

1 dρ
R = R0 [1 + α (T − T0)] with α (T ) = , (3.3.25)
ρ dT

where α is the temperature coe cient and ρ is the resistivity of the material.
A typical temperature coe cient for metal is α = 0.005 K-1, while for semiconductors α is negative
and an order of magnitude higher: α = -0.06 K-1. The large temperature coe cient of semiconductors
can be easily understood considering the increase in carrier density with T:

(Ec − Ev)
n p = Nc Nv e− KT (3.3.26)

Figure 3.14 shows the typical dependance of resistivity vs. temperature for di erent materials. We
can observe that in case of a semiconductor active element, the highest detection sensitivity can be
achieved when it is cooled to low temperature, where the steepness of the R-T characteristic is max-
imum. The lowest limit is given by 4.2 K (Liquid helium temperature) corresponding to an extremely

Figure 3.14 Behaviour of the resistance vs. tempera- Figure 3.15 Typical operating circuit for a bolometer.
ture for different materials.

high temperature coe cient. Alternatively, high sensitivity can be reached using high temperature
superconductors (e.g. YBa2Cu3O7 - Tc = 95K) operated close to the critical temperature. Figure 3.14.
The change in the resistance of the absorbing element can be measured with an active circuit, de-
signed to minimise heating by dissipative e ects (high input impedance).
In composite bolometers the absorbing element is separated by the thermometer. The radiation
absorber is deposited on a substrate connected to a heat sink by a heat conducting path. Radiation
energy is converted into heat, which raises the temperature of the absorber. The active element of the
thermometer is in thermal equilibrium with the absorber so that the temperature increase induces a
change in its resistance, which is detected by measuring the electrical signal across it. Figure 3.16
shows a composite bolometer designed to detect THz radiation. The absorber is made by a thin layer
of bismuth deposited onto a diamond substrate facing a thermometric unit based on a silicon doped
resistance. Bismuth is widely used for the absorbing material because of its relatively low thermal
conductivity, high optical absorption, and very small speci c heat. Diamond is considered to be the
best material for the substrate because of its excellent thermal conductivity and high transparency in
an extremely wide spectral range that includes the THz region. Heavily doped Si or Ge are the most
widely used materials for the thermometer. The doping brings the material close to the metal-insula-

89
ffi

ffi
ffi
ff

fi

ffi

ff

Figure 3.16 Composite bolometer designed to detect


terahertz radiation.

tor transition so that the electrical resistance is low enough at cryogenic temperature and sensitive to
minute temperature changes.
A new emerging technology based on the bolometer concept allows the production of relatively
low cost thermal imagers operating in the Long Wavelength Infrared (LWIR) band (7-14 μm), which
corresponds to the wavelength of the radiation emitted by living bodies, like humans and animals.
Beyond application in security, remote sensing and other civil and industrial elds, those devices,
called micro-bolometers, could become popular in a near future in the automotive industry. In fact,
few car manufacturers have already introduced thermal imagers in high-end models to provide the
driver with visual alerts for pedestrians or animals crossing the road in complete darkness, as shown
in (Figure 3.17).

Figure 3.17 Thermal imager based on a micro-bolometer array

A micro-bolometer is made by a matrix of IR absorbing elements (pixels), which also constitute


the electric resistances. Each pixel has its own reading circuitry (Figure 3.18), and arrays of 640x480
or 1024x768 pixels are nowadays available. The most used materials for absorbing elements are
amorphous silicon and vanadium oxide. Even if micro-bolometers are less sensitive than cooled IR
quantum detector, they work at room temperature, which makes them suitable for the consumer
market. The performance is expressed by the Noise-Equivalent Temperature Di erence (NETD),
which is an alternative expression for NEP. A typical merit gure is NETD ≈ 50 mK or less, which
makes micro-bolometers sensitive enough for thermal imaging.

90

fi

fi
ff
Figure 3.18 Picture element of a micro-bolometer array and
schematic of the readout circuit.

Photoacoustic detectors

THE GOLAY CELL

The Golay cell is a pneumatic detector rst proposed by Marcel J. E. Golay in 1947. In Golay cells
the absorption of optical radiation induces the thermal expansion of a gas. The radiation absorber of a
Golay cell is typically a blackened metal lm on a thin substrate. The heat is transferred to a small
volume of gas in a sealed chamber behind the absorber so that the pressure in the chamber increases.
A re ective and exible membrane is attached to the back side of the chamber, and an optical lever-
age detects the membrane deformation induced by the pressure increase.
Figure 3.19 shows the basic components and operating principle of a typical Golay cell designed
for terahertz detection. Modulated THz radiation passes through the front window and is collected by
the absorbing lm. The released energy heats up a small volume of gas enclosed in the pneumatic
chamber, and induces thermal expansion of the gas. The resulting pressure rise deforms the exible
mirror attached to the backside of the chamber. An optical beam from a light emitting diode is fo-
cused on the exible mirror. The re ected beam is collected and focused onto a photodetector. The
de ection caused by the membrane deformation is sensed by the detection readout system, which is
usually based on a four quadrant photodiode. The unbalance in the signal current exiting the photo-
diodes allows one to calculate the change in the radius of curvature of the membrane and, in turn,
the power released inside the cell.

Figure 3.19 Schematic of a Golay cell for detecting THz radiation.

91
fl
fl
fl
fi
fl

fl
fi
fi

fl
High detection sensitivity of a Golay cell requires several conditions for its components. Ideally,
the absorbing lm should be the only medium experiencing heat exchange. To achieve this result,
materials for the window and the pneumatic chamber must be heat insulators of high quality, and the
gas should be transparent over the whole detection spectrum. The heat conductivity of the gas should
also be small. Xenon is commonly used, because it is the least conductive gas.

Figure 3.20 Golay cell made by Microtech Instru-


ments (http://www.mtinstruments.com)

The Golay cell is the most sensitive detector among thermal radiation detectors which operates at
room temperature. Responsivity of a Golay cell is in the range of kV/W when the modulation fre-
quency is a few tens of hertz. Typically obtained NEP is 0.1-1 nW·Hz1/2.
The main properties of the Golay cell shown in Figure 3.20 are listed in Table 3.1.

Table 3.1. Main parameters of the Golay cell shown in Figure 3.17

PARAMETER VALUE

Responsivity at 12.5 Hz modulation 104 V/W

Sensitivity at 12.5 Hz modulation 10-8 W/H1/2

Maximum modulation frequency 50 Hz

Dynamic range 100 nW -1 mW

Rise time 25 ms

Maximum output voltage 3.0V

Input window diameter 6mm

Dimensions 135 x 115 x 120 mm

Pyroelectric detectors
Pyroelectric detectors rely on crystals that show electric polarisation susceptible to changing with
temperature. Pyroelectricity is a rather common property of crystals. Yet, detectors are usually made

92
fi

with ferroelectric materials, which exhibit a spontaneous electric dipole moment even in the absence
of an external electric eld. All ferroelectric materials are pyroelectric, but not vice-versa. In the fer-
roelectric state, the center of positive charge in the unitary cell does not coincide with the center of
negative charge. The plot of polarisation versus electric eld for the ferroelectric state shows a hys-
teresis loop, like in ferromagnetism. Hence the name of these materials for analogy with the be-
haviour of iron alloys. Ferroelectricity usually disappears above a certain temperature called the tran-
sition temperature or Curie temperature. Above the transition the crystal is said to be in a paraelec-
tric state. The term paraelectric suggests an analogy with paramagnetism: close to the transition tem-
perature there is usually a rapid drop in the dielectric constant as the temperature increases. In some
crystals, the ferroelectric dipole moment is not reversed by an electric eld of the maximum intensity
which it is possible to apply before causing electrical breakdown.
Ferroelectric crystals undergo a variation in the spontaneous polarisation when the temperature is
changed, making them ideal for infrared light detection. Ferroelectric materials include ionic crystal
structures closely related to the perovskite and ilmenite structures. As an example, we can consider
barium titanate, whose crystal structure is shown in Figure 3.21.
Above the Curie temperature the ferroelectric barium titanate (BaTiO3) resembles calcium titanate
(perovskite), which is the prototype crystal (Figure 3.21 a). The structure is cubic, with Ba2+ ions at
the cube corners, O2- ions at the face centres, and a Ti4+ ion at the body center. Below the Curie tem-
perature the structure is slightly deformed, with Ba2+ and Ti4+ ions displaced relative to the O2+ ions,
thereby developing a permanent dipole moment (Figure 3.21 b).
Commonly used materials for pyroelectric detectors are triglycine sulfate (TGS), deuterated

Figure 3.21 Structure of the prototype crystal of perovskite a)


compared to the structure of the ferroelectric barium titanate
b).

triglycine sulfate (DTGS), lithium tantalate (LiTaO3), and barium titanate (BaTiO3).
The schematic of a pyroelectric detector is shown in Figure 3.22. A thin slab of ferroelectric crys-
tal suitably poled and oriented is sandwiched between electrodes, which are connected to an ampli er
with high input impedance (Figure 3.23). At thermal equilibrium, the spontaneous polarisation vec-
tor P causes the presence of a bipolar polarisation charge of density ±σp. Since the equilibrium of
matter requires global charge balance, the charge densities ±σp must be compensated by equal and
opposite densities of free charge ±σ on the conductive electrodes. This makes the detector equivalent
to a capacitor that stores a charge Q:

| Q | = A | σ | = A | σp | with σp = P ⋅ u n, (3.3.27)

where A is the area of the pyroelectric slab and un is the normal to its surface.

93
fi


fi

fi

fi
Figure 3.22 Schematic of a pyroelectric detector.

Figure 3.23 Simpli ed equivalent circuit of a pyro-


electric detector.

The increase in temperature, upon absorption of IR radiation, induces a variation of the polarisa-
tion charge | σp | that in turn a ects the free charge Q. This results in a current i(t) given by the time
derivative of Q:

dQ(t) d σ (T (t))
i(t) = =A , (3.3.28)
dt dt
where the dependence of σ on the temperature has been explicitly expressed.
The response of the material to a change in temperature is described by the pyroelectric coe cient
p (Table 3.2), which is de ned by the derivative of the polarisation vector P with respect to the tem-
perature:

d P
p= . (3.3.29)
dT
Hence, from Eqs. (3.2.7), (3.2.8) and (3.2.9) the output current that results from an incoming optical
radiation is proportional to the pyroelectric coe cient and to the derivative of temperature:

d σ (T (t)) d σ (T (t)) d T (t) d T (t)


i(t) = A =A ⋅ =pA . (3.3.30)
dt dT dt dt

94

fi
fi

ff
ffi

ffi
From the general theory of thermal detectors (see Eq. 3.3.4), when the detector receives a sinu-
soidally modulated optical power P(t) = W0 + Wm sin(2π f t), the time derivative of its temperature
is given by:

d T (t) dΔT (t) 2π f Cth


= = Wm cos(2 π f t) with τth = . (3.3.31)
dt dt Gth 1+ 4π 2 f 2 τth
2 Gth

where Cth is the thermal capacitance of the pyroelectric material.


Therefore, the current responsivity RI of a pyroelectric detector, for a modulated input, is given by:

i RMS 2π f
RIRMS = = pA . (3.3.32)
Pin
RMS
Gth 1+ 4π 2 f 2 τth
2

When the frequency f is much higher than the inverse of the thermal time constant τth Eq. (3.3.32)
simpli es to:

pA pA
RIRMS = = , (3.3.33)
Gth τth Cth

From Eq. (3.3.32) we can infer that pyroelectric detectors do not respond to CW optical signals
due to the derivative in Eq. (3.3.30) that determines the zero at origin of the frequency axis for the
responsivity (see Figure 3.24). When the modulation frequency is larger than the thermal pole, the
current response is almost at and proportional to the ratio between the pyroelectric coe cient and
the heat capacitance of the detector.
To calculate the voltage responsivity we must consider the equivalent circuit of a pyroelectric de-
tector, which is shown in Figure 3.23. The circuit includes: i) the signal generator i(t), which ac-
counts for the current responsivity RI; ii) the capacitance Ce of the detector, which is far from being
negligible and iii) the shunt resistor Rd, representing the resistance of the ferroelectric slab, which is
typically very high (1011 Ω) since ferroelectric materials are good insulators. Rd can be neglected
compared to RL. The signal is picked by by a FET transistor used as voltage follower that transforms
the very high impedance output V(t) from the detector into a low impedance voltage given by:
V(t) = RL⋅i(t).
The voltage responsivity RV can be calculated from Eq. (3.3.32) considering the network made by
the capacitance Ce in the equivalent circuit and the load resistance RL:

pA τRC τth
RVRMS = 2π f , (3.3.34)
Cth Ce 1 + 4π 2 f 2 τRC
2 1 + 4π 2 f 2 τth
2

where τe = RL·Ce is the time constant of the electric circuit.


From Eq. (3.3.34) it can be observed that a pyroelectric detection system has a transfer function
characterised by 1 zero at the origin and 2 poles, due to the energy storage elements made by the
thermal capacitance Cth and the electric capacitance Ce.
The thermal pole is typically at a frequency from few Hz to tens of Hz, while the electric pole can
be set by the choice of the load resistance RL. High speed systems can be obtained only with a low
value of RL, which leads to a severe reduction in the responsivity, being the gain-bandwidth product
of the system constant, as shown in Figure 3.24.

95
fi


fl

ffi

Figure 3.24 Responsivity of a detection system


based on a pyroelectric device.

In the frequency interval between the thermal pole (1st) and the electric one (2nd), where the de-
tector is typically used, Eq. (3.3.34) can be simpli ed to:

p A τRC p A RL
RV = = . (3.3.34a)
Cth Ce Cth

In case the electric pole is at frequency lower than the thermal pole, which is the case in high respon-
sivity detectors operating at low frequency, the responsivity can be approximated to:

p A τth pA
RV = = (3.3.34b)
Cth Ce Gth Ce

Equations (3.3.34a) and (3.3.34b) highlight the proportionality of the response to the load resis-
tance, as shown in Figure 3.24.
Finally, we observe that the volume of the pyroelectric slab is given by its area A times the thick-
ness d. By calling cth the volume speci c heat of the ferroelectric material, we can derive that the cur-
rent responsivity RI, in the frequency region beyond the thermal pole, is proportional to:

p
RI ∝ , (3.3.35)
d cth

which would suggest the choice of a thin slab. Yet, the thickness of the slab must take into account
the absorption coe cient of the material and its reduction a ects negatively the bandwidth of the
voltage responsivity, which is the ultimate parameter for practical applications. In fact, the useful
bandwidth, which is set by τRC, is proportional to:

d
fRC ∝ . (3.3.36)
ϵr

96


ffi


fi
fi

ff

In the end, the design of a pyroelectric detector should privilege those materials featuring high
pyroelectric coe cient, low relative permittivity and low speci c heat. These parameters are listed in
Table 3.2.

Table 3.2 Properties of ferroelectric materials used in py-


roelectric detectors.

With a proper design of the readout circuit, the responsivity can be up to 103 V/W and the NEP
can be as low as 10-9 W/Hz1/2.
High end pyroelectric detectors are used in Fourier Transform Infrared spectrometer (FTIR), in
remote sensing and for terahertz detection. Yet, the most common application of pyroelectric detec-
tors is in anti-intrusion systems, where low cost detectors provide enough sensitivity to sense the IR
radiation emitted by the human body. For this application the absence of CW response is not a con-
cern. Further, the frequency response of these detectors matches rather well the typical time variation
of the IR signal emitted by a person wandering in a room.
Finally, arrays of pyroelectric detectors can be manufactured by assembling a large number of ele-
ments similar to the one shown in Figure 3.25, in such a way to build an e cient and relatively low
cost thermal imager.

Figure 3.25 Micro-element of a pyroelectric array.

97

ffi

fi
ffi
Section 4

QUANTUM DETECTORS
Quantum detectors measure the photon ux impinging onto the sensitive area. They di er from
thermal detectors, which measure the power of the incident radiation eld, because their output de-
pends on the spectral content of the incoming radiation. In the present section we will consider only
those detectors characterised by a linear relationship between the number of absorbed photons and
the output signals, which is primarily made by a current. In high energy physics, from X-ray on, the
detection process takes place through multiple quantum interaction per each photon or particle, lead-
ing to an output signal that accounts for the energy of the incoming photons and not just for their
number. Detectors for high energy radiations are beyond the scope of the present discussion.
In quantum detectors, photons in the UV - IR spectral region interact with the valence electrons
inside the active material, which is typically a semiconductor, and might induce quantum transitions
that result in an output current. In many detectors the signal is simply given by this current, eventu-
ally converted into voltage by a transimpedance ampli er. The signal is then proportional to the pho-
ton ux, through a parameter called quantum e ciency, which describes the likelihood of collecting
an output electron per absorbed photon. The quantum e ciency is function the wavelength.
When the light signal is very weak, noise prevents us to measure the current solely given by the
conversion of the incoming photons into electrons, which might correspond to a current of few elec-
trons per second. In these cases, ampli ed detectors can be conveniently employed. Yet, it is worth
noting that, in order for ampli cation to be e ective, it must take place as soon as possible, which
means inside the detector. In fact, external ampli ers rarely succeed in extracting a signal buried
within noise. Several ampli cation mechanisms, based on electron multiplication within the detector,
have been devised to greatly increase the sensitivity of quantum devices that, in the best operating
conditions, can reach single photon sensitivity. In the next sections we will see a variety of quantum
interactions that have been exploited to build photodetectors and we will discover that, for the most
important detector types, an ampli cation mechanism has also been implemented to increase their
sensitivity. Nevertheless, a warning is necessary not to create any misconception. Is ampli cation cost
free? Certainly not, and by saying this we are not considering only the increase in complexity and
monetary cost! Being ampli cation the result of a multitude of stochastic events, some extra noise is
always added to the intrinsic noise of the incoming signal. This drawback is measured by a parameter
called Noise Figure, which characterises any ampli er. The Noise Figure is the ratio between the SNR
before and after the ampli cation process, usually expressed in logarithmic units (dB). Hence, ampli-
cation should be used with parsimony, only when the increase in the electron ux at the detector
output makes negligible the noise sources coming later in the detection process, like the Johnson
noise discussed in Section 2.2. In fact, in low light condition the Johnson noise often overcomes the
signal if no ampli cation occurs.
Before ending this short introduction to quantum detectors, I would like to spend few words to
illustrate how ampli cation in light detection is not only the result of the technological revolution of
the modern era, but is something very ancient.
In fact, considering the human eye, a biochemical ampli cation mechanism makes scotopic vision
so e ective to let us see in almost complete darkness, and some animals can do even better. This is
certainly a gift of the Darwinian evolution that took place in the retina (section 1.7). When rhodopsin
absorbs a photon its 11-cis-retinal chromophore rapidly isomerises, causing the cytoplasmic surface

98
fi
ff
fl
fi
fi
fi
fi
fi
fi

fi
fi
fl
ff

ffi
fi
fi
fi
ffi
fi
fi
fl
fi
ff

of this protein to become catalytically active. In this state, Rhodopsin activates the Guanosine
Triphosphate (GTP) binding protein, called Trasducin (T). Within a fraction of a second a single ac-
tive Rhodopsin causes hundreds of Transducin proteins to exchange Guanosine Diphosphate (GDP)
for GTP, forming active TGTP complexes. A greatly ampli ed signal now passes to a third protein
(cGMP PDE), which is activated by TGTP and continue the transduction process. Hence, a single
photon causing the isomerisation of 11-cis-retinal results in a rather strong signal passed to the
transduction chain that involves the optical nerve and terminates in the brain cortex.
The photographic lm, which has allowed your fathers and grandfathers to take pictures with a
relatively short exposure time, is another clear example of signal ampli cation in image acquisition.
The photographic emulsion is made of a dispersion of very small grains of silver halides (AgI, AgBr,
AgCl) within a gelatin lm. The latent image is an invisible pattern produced by the exposure to light
of the photographic emulsion. Few photons are required to activate a single grain that contains a
huge number of unit cells of silver halide crystals. Then, the development process chemically reduces
the whole activated crystals of silver halide to metallic silver, leading to a negative image. As a matter
of fact, the number of silver ions reduced during the development phase is much greater than the
number of absorbed photons. This constitutes a large ampli cation factor that has made feasible the
photographic process since the second half of the 19th century. The mechanism has been improved by
lm manufacturers up to the end of the past century, by controlling the chemistry of the emulsion
and the size of silver halide grains. Thereafter, image recording has changes dramatically with the
advent of digital photography. We will see in the last section of these notes that the ultimate sensitiv-
ity in imaging devices is once again achieved through on-chip ampli cation.

Photoemissive detectors

PHOTOELECTRIC EFFECT

The photoelectric e ect is well known for being a milestone in the history of physics. In 1921 Al-
bert Einstein was awarded with the Nobel Prize for the explanation of the photoelectric e ect (1905)
that paved the way to the discovery of the laws of quantum mechanics. Beyond this historical per-
spective, the emission of electrons by materials upon absorption of photons is the starting event in a
large category of photodetectors, from vacuum photodiodes to photomultiplier tubes and image in-
tensi er.
When a radiation of wavelength λ impinges on a metal, electron emission takes place provided that
the photon energy is greater than the work function Φ of the metal. The work function is the energy
di erence between the fermi level and the vacuum level (Figure 3.26), which is the energy of an
electron at rest (zero kinetic energy) far away from any material. Hence, the work function is the en-
ergy that electrons must acquire from the electromagnetic eld or any other source of energy, to be-
come free particles. This sets a wavelength threshold λth for electron emission, given by:

hc
λth < . (3.4.1)

When photons with wavelength shorter than λth are absorbed by the metal, the energy in excess be-
yond the work function might become kinetic energy of emitted electrons. Yet, in general most of the
kinetic energy is quickly dissipated in collisions with thermal electrons and lattice ions. In particular,
the collision of a fast electron with a thermal one leads to a large energy loss because the two parti-
cles have the same mass. Therefore, only a small fractions of electrons that do not undergo any colli-
sion can escape the metal with the maximum kinetic energy:

Ek ma x = h ν − Φ (3.4.2)

99
fi
ff
fi



fi
ff
fi
fi
fi

fi
fi
fi

ff
Because of collisions, most of the excited electrons exit the metal with a kinetic energy lower than
the maximum or do not exit at all, being eventually thermalised after few interactions with other
electrons.
Photoemissive detectors rely on electrons emitted, upon irradiation, by a photocathode, which is
made by thin layer of a sensitive material kept at negative potential within a vacuum tube. Di erent
types of photocathodes have been devised with a wide range of properties in terms of spectral sensi-
tivity, quantum e ciency and dark current (i.e. noise).

Figure 3.26 Scheme of energy levels for electron


emission by metals.

PHOTOCATHODE MATERIALS

Photocathodes can be theoretically made by metals and this was the case at the time of the rst
experiments. Yet, metals are unsatisfactory in terms of spectral sensitivity and quantum e ciency.
Cesium is the alkali metal with the lowest work function (Φ = 1.95 eV). This sets the threshold
wavelength for electron emission to λth ≅ 0.6 μm, which is far from optimal. Further, the quantum
e ciency η of cesium, that is the ratio between the number of emitted electrons to the number of
absorbed photons, is really poor: η ≅ 0.1%. In fact, due to the large density of free carriers in metals,
the thermalisation of fast electrons takes place in a very short time, before they have the chance to
get to the surface. This makes photoemission a rare event.
For the above reasons, modern photocathodes are made by a thin layer of intrinsic or weakly
doped semiconductor deposited over a transparent substrate, which might be borosilicate glass or UV
transparent materials, like MgF2. Electrons can be emitted either from the same side that receives the
radiation or from the opposite side, being the latter the most common option. In this case, the pho-
tocathode thickness is a compromise aiming at maximising the probability that a photon produces an
emitted electron. If the layer is too thin, little light can be absorbed; if it is too thick, the resulting
photoelectrons cannot e ciently escape from the photocathode.
The Lambert-Beer law, which describes the exponential attenuation of a collimated light beam
within a material, can be used to set the proper thickness of the photocathode:

I(x) = I0 e −μa x, (3.4.3)

where I(x) is the intensity of the beam as a function of the position within the material, I0 is the in-
coming intensity and μa is the attenuation coe cient. The inverse of the attenuation coe cient, i.e.
the penetration depth L = 1/μa provides a rough estimate of the required thickness. As a matter of

100
ffi


ffi
ffi
ffi

ffi
ffi
fi
ff
fact, the con icting requirements of high photon absorption and e cient electron escape lead to
cathodes with a thickness of a few tens of nanometers.
Conversely, with metals, free electrons are rare in semiconductor. Hence, the collision of excited
electrons mainly takes place with the ion lattice, resulting in phonon exchange. Since the energy of
phonons is much lower than the kinetic energy of excited electrons, the energy loss per single event
is low and thermalisation requires many interactions, which makes the quantum e ciency of semi-
conductor photocathodes much higher than that of metals. A further advantage of semiconductors is
the low re ectivity, which helps increasing the e ciency.
To describe the spectral sensitivity of semiconductor photocathodes, the work function Φ is not
appropriate, since the Fermi level falls within the band gap, a region forbidden for electrons. In fact,
for an electron to be ejected from the surface of a semiconductor, the energy acquired from the e.m.
eld must be greater than the energy gap Eg plus the electron a nity χ, which is the energy di erence
between the vacuum level and the conduction band (Figure 3.27).
This sets the wavelength threshold λth to:

hc
λth < (3.4.4)
e(Eg + χ)

Figure 3.27 Schematic of the energy levels in an alkali


photocathode.

Photocathodes made of antimony combined with alkali metals are amongst the most commonly
used. More speci cally, Bialkali photocathodes (Na2KSb) are suitable for the visible range, being blind
in the infrared. Better performances can be achieved with Multialkali photocathodes, like NaKCsSb
(also known as S-20), which is sensitive up to near-infrared and is still rather common in many de-
tectors. Finally, a very old photocathode (developed in 1929), still in use, is made of AgCsO (S1). S1
photocathode has an impressive sensitivity extended to 1200 nm, but it is a ected by high dark cur-
rent, i.e. spontaneous emission of electrons by thermionic excitation, and requires cooling.
To improve the spectral sensitivity and the quantum e ciency, Negative Electron A nity (NEA)
photocathodes have been developed. NEA photocathodes are made of a p-doped semiconductor, for
example GaAs, activated by forming a layer of cesium or cesium oxide (Cs2O) on the surface of the
crystal. Cesium is an alkali metal prone to loose electron and to assume a positive charge. Hence, this
deposition behaves as an n-type material resulting in a layer of ionised acceptors at the very surface

101
fi
fl
fl

fi

ffi

ffi

ffi

ffi
ff
ffi
ffi

ff
of the p-type semiconductor, since holes undergo electrostatic repulsion. The charge depletion layer
causes an electric eld directed inward that curves downward the energy bands close to the surface,
similarly to what happens in a p-n junction. The vacuum level is also pulled downward and the elec-
tron a nity diminishes, becoming negative (Figure 3.28). Sometimes, cesium oxide is deposited
onto the surface, through a further oxidation step, as it is shown in Figure 3.28. It might be noted
that Cs2O exhibits a strong dipole moment, which further lowers the vacuum level, thus reducing the
electron a nity χ and the e ective work function Φ. In NEA photocathodes, once electrons have been
promoted from the valence band to the conduction band by photon absorption, they are allowed to
escape even after thermalisation, until recombination takes place, since the bottom of the conduction
band is above the vacuum level. In the case of GaAs:Cs0, the sensitivity is extended to 900 nm and
the quantum e ciency increases up to 25%.

Figure 3.28 NEA GaAs photocathode made with a deposi-


tion of cesium oxide on the surface of the semiconductor.

In this case, the threshold wavelength is given by:

hc
λ= (3.4.5)
e EG

which only accounts for the Energy gap EG.


Other NEA photocathodes are the ones made with InGaAs(Cs) and GaAsP(Cs). The last provides
very high quantum e ciency (up to 50% within a speci c wavelength range) even if its spectral sen-
sitivity is limited to 750 nm.
Figure 3.29 shows the radiant spectral sensitivity of common photocathodes, as a function of the
wavelength.
Table 3.3 illustrates the most important speci cations of some common alkali and NEA photo-
cathodes. Radiant sensitivity is de ned as the photocathode current emitted per unit power (watt) of
incident radiation at wavelength λ and is expressed in mA/W. It is related to quantum e ciency η,
which is represented in Figure 3.29 by dashed lines. Lines of equal quantum e ciency raise with
wavelength to take into account the decrease in photon energy. The luminous sensitivity indicates the
current emitted per unit of luminous ux (lumen) in the incident radiation. The most common mate-
rials for transparent windows are also reported in Table 3.3: Borosilicate glass is used for general
purpose, quartz is used for UV down to 300 nm. MgF2 is required for deep UV.

102
ffi
ffi

ffi
fi
ffi
ff

fi
fl

fi

fi

ffi

ffi
Figure 3.29 Radiant spectral sensitivity of common photocathodes as a func-
tion off the wavelength (source hamamatsu Photonics).

Table 3.3 Properties of common photocathode materials.

VACUUM PHOTODIODES

The vacuum photodiode is the simplest photoemissive detectors. It is is made by a a glass tube
with a photocathode, put inside the tube or deposited to the inner surface of the tube, and an anode.
A high voltage power supply sets a potential di erence ΔV ≅ 100-200 V between the anode and the
cathode. The resulting electric eld is required to collect at the anode all the electrons emitted by the
photocathode, upon illumination. The schematic of a vacuum photodiode is shown in Figure 3.30.

103
fi
ff

The output current as a function of the supply voltage for di erent values of the incoming photon
ux Φ is shown in gure Figure 3.31. The device is typically operated in the active region, where the
characteristics are almost at and the response is linear to the ux.
A small positive slope of the characteristics is due to the presence of the bias electric eld, close to
the photocathode, that bends downward the vacuum level and slightly diminishes the work function,
as shown in Figure 3.32. Yet, this e ect can be often neglected and the vacuum photodiode can be
assumed equivalent to an ideal current generator controlled by the photon ux Φλ.
The output current is given by:

Pλ ηeλ Pλ P λ
Iλ = Φλ ηe = Φλ = = λ , (3.4.6)
hc hν hc
where Pλ is the incoming optical power at wavelength λ.
The responsivity, in terms of current or voltage across the load resistor RL is then:

Iλ λ Vλ λ
Rλ(I ) = = ηe Rλ(V ) = = ηeRL (3.4.7)
Pλ hc Pλ hc

Vacuum photodiodes have low sensitivity, fast temporal response and typically require external
ampli cation. Hence, nowadays they are seldom used, unless for speci c application involving UV
light detection, when silicon photodiodes can be hardly employed. They can be used to detect light
pulses and in analytical instruments.

Figure 3.30 Schematic of a vacuum photodiode. Figure 3.31 Characteristics of a vacuum photodi-
ode for different values of photon ux. The satura-
tion region is marked.

Noise in Vacuum Photodiodes


An intrinsic noise source in vacuum photodiodes, which is present in all the photoemissive de-
vices, is given by the dark current of the photocathode. Dark current is due to thermionic emission. A
simple model of thermionic emission is given by the Richardson-Dushman equation, that expresses
the current density JT exiting a metal as a function of the temperature T and the work function Φ.

104
fl
fi
fi
fl

fl

ff

ff
fl

fi
fl

fi

( kT)
k 2 me e k 2 me e
[ 2π ℏ ]
Φ
JT = T 2 ex p − α = , (3.4.8)
2 3 2π 2ℏ3

( kT)
Φ
JT = α T 2 ex p − . (3.4.9)

The term within square parentheses in Eq. 3.4.8 depends on fundamental quantities and is a constant
approximatively given by α ≅ 1.2⋅106 A m-2 K-2. For semiconductor photocathodes the constant α must
be multiplied by a correction factor that depends the the material the cathode is made of. As a rule of
thumb, typical values of the dark current density @300 K, for photocathodes are in the range
10-15-10-11 A⋅m-2. The dark current of a photocathode is proportional to its area A. From Eq. (3.4.9), it
is evident that cooling is very e ective in reducing the dark current, but below a certain temperature,
cooling might also decrease quantum e ciency. Further, other noise sources that do not depend on
temperature a ect the electron emission by the photocathode. Among them, we might consider the
contribution of high energy photons, like cosmic rays, and the so called ionic feedback. In fact, even
in high vacuum, few ions are always present in the tube. Positive ions are attracted by the cathode
and, upon collision, induce the emission of a bunch of electrons for impact ionisation.
The second important noise source in a vacuum photodiode is the shot noise (Section 2.3). In fact
the output current originates from a stream of electrons collected at the anode, thus leading to ran-
dom uctuations in iRMS that intrinsically depend on the quantisation of the electric charge. To make
things more clear, let us consider only the shot noise which stems from the signal current, even if we
known that also dark current contributes to shot noise. In low light condition, assuming a stationary
photon ux Φs, the output current can be represented by the plot shown in Figure 3.33, where the
random uctuation of the current are due to the granularity of the charge ow.
The average signal current Is is proportional to the input power Pin and to the wavelength λ

e
Is = Rλ Pin = η λ Pin. (3.4.10)
hc
From the theory of shot noise, the root mean square noise current is given by:

⟨(δi ) ⟩ = ⟨[i(t) − Is] ⟩ =


RMS 2 2
iShot = 2eIsΔ f , (3.4.11)

where Δf is the bandwidth of the detector. The Signal to Noise Ratio (SNR) is then given by:

Figure 3.32 Effect of supply voltage on electron Figure 3.33 Signal and noise in a vacuum photodiode
af nity.

105
fi
fl
fl
fl
ff


ff
ffi


fl

S Is Is2 Is Pin ηλ
= = = = . (3.4.12)
N 2eIsΔ f 2eΔ f 2h cΔ f
⟨(δi ) ⟩
2

A more common de nition of SNR involves quadratic quantities, such as it is the ratio of the power
of a signal to the power of noise:

S Is2 Pin ηλ
= = . (3.4.12a)
⟨(δi ) ⟩
N 2 2h cΔ f

This de nition has the advantage of being independent on the sign of the signal Is, which can be ei-
ther positive or negative. Finally, the SNR is often expressed in the logarithmic decibel (db) scale:

(N) (N) ( 2h cΔ f )
S S Pin ηλ
= 10 log10 = 10 log10 . (3.4.12b)
dB

It is worth noting that the SNR, for the same value of the input power Pin, increases with the wave-
length λ and decreases with the bandwidth of the detector Δf. This can be easily understood consider-
ing that in quantum detectors the signal is proportional to the photon ux, while the e ect of the
bandwidth on SNR derives from the white model that has been assumed for the shot noise.
As a pure academic exercise, from Eq. (3.4.12), which only considers the shot noise of the sig-
nal, we can calculate the very limit for the power of a detectable radiation, by setting SNR = 1.

2h cΔ f
Imin = 2eΔ f Pmin = (3.4.13)
ηλ

Since in very low light condition a large bandwidth is not required, we can assume Δf = 1 Hz. The
wavelength can be set to λ = 400 nm, while the quantum e ciency could be η = 10%. This gives: Imin
≈ 2 e-/s and Pmin ≈ 10-17 W (20 ph./s). Beware that this gure does not represent the NEP of the de-
vice because the most relevant noise sources have been neglected. In fact, this minimum power is far
from what can be measured with a real device, mainly due to the Johnson noise considered thereafter.
At any rate, this exercise tells us that the shot noise does not limit the sensitivity of vacuum photodi-
odes, and this is often the case for quantum detectors in general applications.
For a more realistic estimation of the sensitivity of a photodiode, rst of all, we need to replace the
signal current Is with the total current IT in order to calculate the shot noise:

IT = Is + Id + Ib, (3.4.14)

where Id is the dark current due to thermionic emission, while Ib is the background current due to
cosmic rays and other e ects like ionic feedback. Yet, including all the contributes to the current does
not change much the scenario, since the most critical noise source in a photodiode is the thermal
noise of the load resistor RL, which is included in the schematic of Figure 3.30:

4k T
ithRMS = Δ f. (3.4.15)
RL

This noise source, in fact, is the one that sets the noise equivalent power (NEP) of vacuum photodi-
odes, as we will see in a minute. Without any input radiation the output current in is solely given by
noise sources:

106
fi



fi
ff




fi
ffi
fi
fl

ff

( RL )
4k T
inRMS = 2eId + Δ f, (3.4.16)

where only the thermionic dark current Id has been considered to calculate the shot noise.
To nd the NEP, we must calculate the power of a sinusoidally modulated light Pin(t) that, in a noise
free detector, would give a RMS signal current is equal to in given by Eq. (3.4.16).

Pin(t) = P0 (1 + si n ωt). (3.4.17)

From the Responsivity R, the expected RMS value of the signal current is:

P0 P0 e ηλ
isRMS = P RMS R = R= , (3.4.18)
2 2 hc

The NEP can be nally calculated by setting isRMS = inRMS, which corresponds to SNR = 1 as shown
in Figure 3.34. The minimum detectable optical power in modulated regime is then:

SNR=1

( RL )
inRMS h c RMS hc 4k T
N EP = P RMS = = in = 2eId + Δ f. (3.4.19)
R e ηλ e ηλ

To estimate the value of NEP for a vacuum photodiode, we can assign typical values to the parame-
ters in Eq. (3.4.16), which gives the noise current. Assuming a cathode area of few square centime-
tres, a typical dark current @ 300 K is Id = 10-11 -10-12 A. Hence, the value of the rst term between
round parentheses (shot) is about 10-30 A2⋅s. The second term, which accounts for thermal noise, is
10-22 A2⋅s at room temperature, for the standard value of the load resistor RL = 50 Ω. Therefore, it
exceeds the shot noise term by eight orders of magnitude. The thermal noise could be reduced by
increasing RL to higher values, but this would lead to an unacceptable time response, considering the
capacitance of the device. Also cooling can be bene cial, but Johnson noise still prevails. Neglecting
the shot noise of the dark current Eq. (3.4.19) becomes:

Figure 3.34. Comparison between signal and noise when SNR=1,


as required to calculate the NEP.

107
fi
fi



fi

fi

hc 4k T
N EP = Δf (3.4.20)
e ηλ RL

and we get NEP ≅ 1μW @ Δf=1MHz, T=300 K, which prevents the use of vacuum photodiodes in
low light conditions.
In the next subsection we will learn how internal ampli cation in photoemissive detectors greatly
improves NEP, pushing the sensitivity of these devices to the single photon limit.

PHOTOMULTIPLIER TUBES

A photomultiplier tube (PMT) is a photoemissive detector consisting of a light input window, a


photocathode, an electron multiplier, made by a sequence of dynodes, and an anode, assembled into a
vacuum container, as shown in Figure 3.35.

Figure 3.35 Schematic of a photomultiplier tube (source Hama-


matsu Photonics).

Dynodes are made of materials prone to emit secondary electrons upon receiving the kinetic ener-
gy of incoming electrons accelerated by an electric eld. The electric eld is established by a bias net-
work that divides the potential drop generated by an external high voltage supply amongst the cath-
ode (-HV), the dynodes and the anode (ground potential), as shown in Figure 3.36.

Figure 3.36 Resistive bias network of a photomultiplier


tube.

When light enters the window, photoelectrons are emitted by the photocathode and are then ac-
celerated and focused so as to strike on the rst dynode where electron multiplication takes place by

108

fi
fi
fi
fi

secondary emission. This secondary emission is repeated at each of the subsequent dynodes, result-
ing in clusters of 106 - 107 electrons that are collected by the anode.
The dynodes are disposes in di erent arrangements inside PMTs as shown in Figure 3.37. The
design of electrodes must be optimised through an analysis of the electron trajectory in order to
maximise the collection of photoelectrons and secondary electrons by dynodes and also to minimise
the spread in electron transit time. The electron trajectories depends on the dynode shape and on the
applied electric eld. Special care has to be put on the optimisation of the trajectories from the pho-
tocathode to the rst dynode, because the time performance of the PMT mainly depends on the min-
imisation of this transit time.

Figure 3.37 Common con gurations of photomul-


tiplier tubes.

The dispersion of transit times spent by electrons between dynodes is caused by the di erent
emission speed of primary and secondary electrons and the dispersion of electron paths. The best
response (few nanoseconds) is achieved by focused con gurations. Figure 3.38 shows the electron
trajectories in a linear-focused type PMT.
The major secondary emissive materials used for dynodes are alkali antimonide, beryllium oxide
(BeO), magnesium oxide (MgO), gallium phosphide (GaP) and gallium arsenide phosphide (GaAsP).
These materials are coated onto a substrate electrode made of nickel, stainless steel, or copper-beryl-
lium alloy. When a primary electron with energy Ek strikes the surface of a dynode, δ secondary elec-
trons are emitted on average. Figure 3.39 shows the secondary emission ratio δ for various dynode
materials as a function of the accelerating voltage for the primary electrons. The decrease in sec-
ondary emission ratios at high acceleration voltage depends on the increase of electron penetration
depth, which hamper the escape of secondary electrons from the dynode. A typical value of δ for an-
timony based materials is 3-7, while for NEA materials, δ can reach 50 or more. A high secondary
emission ratio is especially important for the rst dynode in order to minimise the multiplication
noise, as we will see in the next subsection. Further, the energy of secondary electrons emitted by
dynodes follows a typical distribution function shown in Figure 3.40. The energy on the x axis is
measured from the Fermi level [E-EF]. The density of electrons start rising beyond the energy gap and
the main peak on the left side asymmetrically distributed around a mean energy of a few electron-
volts, is the spectrum of real secondary electrons. The sharp peak marked with (p) on the right side
represents incoming electrons that are re ected almost without energy loss. The long, low-level line
between the peaks represents the few incoming electrons that undergo multiple inelastic collisions.
The energetic dispersion of secondary electrons a ects the gain and the temporal performances of
photomultipliers.

109
fi
fi

fi

ff

fl
fi
ff
fi

ff
In the simplest case of equal dynodes, the current ampli cation factor or gain G of a photomulti-
plier tube having N dynodes and the average secondary emission ratio δ per stage will be ideally G =
δN. From Figure 3.39, it it possible to observe that, if an appropriate acceleration voltage is set for
primary electrons, a linear relationship exists between the supply voltage VB and δ.

δ = kVB (3.4.21)

G = δ n = k nV Bn,

and the responsivity is:

e
Rλ = G η λ (3.4.22)
hc
Considering a PMT with n dynodes, it is easy to understand that a small variation of the supply volt-
age can seriously a ects the overall gain G:

ΔG n kV Bn−1 ΔVB ΔVB


= =n (3.4.23)
G kVB n VB

Equation (3.4.23) shows that the relative variation of the gain is n times the relative variation of the
supply voltage. Further, when a rather intense light pulse is detected a large current ows toward the
anode due to the ampli cations of photoelectrons. This results in a decrease of the voltage drop on
the resistor between the anode and the last dynode and between the neighbouring dynodes, as it is
shown in Figure 3.41, where the e ect has been exaggerated for clarity. The reduction in the ampli -
cation ratio of the last dynodes adds to the e ect of the space charge given by the multitude of elec-
trons owing toward the anode, which might decrease the charge collected at high current. The men-
tioned e ects could severely in uence the linearity of the response. To stabilise the voltage drop
amongst dynodes, the resistors of the bias network shown in Figure 3.41 can be replaced by zener
diodes for the rst and the last dynodes, which are the most critical for gain stability, as shown in
Figure 3.42.
The time response of a photomultiplier tube can be characterised by delivering onto the photo-
cathode a very short light pulse (few ps) that approximates a delta function. The output current ap-
pears at the anode after a delay of few nanoseconds, required for the ampli cation process to take
place. Due to the already mentioned distribution in the kinetic energies of secondary electrons and
the spread in the electron trajectories, the width of the current pulse exiting the anode is much larger
than the width of the light pulse, as it is schematically shown in Figure 3.43. The time performance
of the detector can by quanti ed by the rise time, which is de ned as the time for the output pulse to

Figure 3.38 Electron trajectories is a photomultiplier tube obtained


through numerical simulation.

110
fl

ff


fi

ff

fi
fi

fl
ff

ff
fi
fi
fi

fl

f
Figure 3.39 Number of secondary electrons emit- Figure 3.40 Energy distribution above the Fermi
ted per primary electron as a function of the accel- level [E–EF] of secondary-electrons.
erating voltage. .

Secondary electron energy (eV)

increase from 10 to 90 percent of the peak height, and by the fall time, which is the time required to
decrease from 90 to 10 percent of the peak height. The position of the pulse peak de nes the electron
transit time from the cathode to the anode.
High gain photomultiplier tubes are capable of detecting single photons. In that case current puls-
es similar to the one depicted in Figure 3.43 appear at the anode, as a result of the ampli cation of a
single photoelectron emitted by the cathode. Then, the dispersion of the peak position (Jitter) mainly
depends on the indetermination of the trajectories of electrons from the cathode to the rst and sec-
ond dynodes. Jitter is particularly critical when the PMT is used in Time Correlated Single Photon
Counting (TCSPC) detection systems. Conversely, the dispersion of the electron paths in the remain-
ing part of the ampli cation chain mainly in uences the width of the output pulses.

Figure 3.41 Voltage drop in the bias network at Figure 3.42 Bias network based n zener diodes to
high anode current. prevent voltage instabilities.

111
fi

fl

fi
fi
fi
Position Sensitive Photomultiplier Tubes
In position sensitive photomultipliers light is captured by a photocathode and converted into pho-
toelectrons as in standard PMTs. Yet, the multiplication process takes place in a 1D or 2D array of
dynodes that work in parallel, as it is shown in Figure 3.44. The bunches of electrons end up in an
array of anodes matched with the multiplication channels (Multianode PMT). In photon counting
applications, the signals collected by each anode, which primarily consists of current pulses, are de-
livered to a Multichannel Analyser. This device counts the pulses keeping trace of the element from
which they were coming from. In this way, it is possible to correlate the counts recorded by the Mul-
tichannel Analyser to the positions where photons were converted into electrons at the photocath-
ode. Typically, position sensitive photomultipliers are coupled to a spectrometer and measure the
spectrum of the incoming radiation.

Figure 3.43 Time response of a photomultiplier Figure 3.44 Multianode position sensitive photo-
tube. multiplier tube

The main advantage of mutianode PMT is their parallel output that is useful when several mea-
surements have to be done at the same time and/or spatial resolution is required. Critical issues are
gain uniformity and crosstalk between channels that limit their performances.
Amongst position sensitive detectors, devices based on micro-channel plate ampli ers deserve a
special mention since they can be used in image intensi ers as much us in high performance photo-
multiplier tubes. Micro-channel plate photomultiplier tubes are the subject of the next subsection.

Micro-Channel Plate Photomultiplier Tubes


The best performances in photon counting can be archived with Micro-Channel Plate (MCP) pho-
tomultiplier tubes, where the discrete dynode set is substituted by an array of miniature electron
multiplier channels, each acting as a continuous dynode chain. The multiplier channels are bundled
in parallel and formed into the shape of a thin disk. This structure exhibits dramatically improved
time resolution as compared to discrete dynode PMTs. It also assures stable gain in high magnetic
elds and provides position-sensitive capabilities when combined with a special anode.
Each channel has an internal diameter ranging from 6 to 20 microns with the inner wall processed
to have high electrical resistance and secondary emissive properties. A voltage of typically 1-2 kV is
applied across the ends of the MCP, producing a weak current that establishes a homogeneous elec-
tric eld inside the channels, which act as independent electron multipliers. The cross section of a
channel and its multiplication principle are illustrated in Figure 3.45. When a primary electron im-
pinges on the inner wall of a channel, secondary electrons are emitted. Being accelerated by the elec-
tric eld these secondary electrons bombard the channel wall again to produce additional secondary

112
fi
fi
fi

fi

fi

electrons. This process is repeated many times along the channel and, as a result of the multiplica-
tion, a large number of electrons are released from the end of the plate.
Amongst the advantages of MCP-PMTs over conventional con guration, we can mention the sim-
ple geometry, which not requires any focusing element and prevents electron losses, the compactness
(typical thickness is 1-2 mm) and the lower dispersion in electron trajectories. The transit time is

Figure 3.45 Micro Channel Plate. Figure 3.46 Common con gurations for multistage
MCPs.

reduced to about 1 ns and the jitter of single photon pulses can be as low as few tens of picoseconds.
Amongst the weaknesses of MCPs we might consider the possible saturation of the anode current
when operated in CW, due to the high resistance of the micro-channels, and the already mentioned
ionic feedback, which is critical due to the high electric eld in microchannels. MCPs are available
also with curved channels which allow a higher gain than normal MCPs without occurrence of ion
feedback because ions do not travel a long path before collision. Preferably, stacked MCPs are em-
ployed to achieve high gain. The channels of each MCP are tilted (8° - 15°) against the MCP normal
and the channels of successive MCPs are tilted to opposite directions. 2 MCPs can be combined in the
chevron con guration and 3 MCPs can be assembled in the Z con guration, as shown in Figure 3.46.

HYBRID PHOTOMULTIPLIERS

In hybrid photomultipliers, also called hybrid photodetectors (HPD) the primary conversion of
photons into electrons takes place in a photocathode, as in standard PMTs, but the electron multipli-
cation is made by a solid state device (Figure 3.47). This consists of a an avalanche diode (see sec-
tion 4.3), where the multiplication, still driven by a strong electric eld, takes place in a semiconduc-
tor. The HPD operates on the following principle: when photons impinge onto the photocathode,
photoelectrons are emitted in proportion to the quantum e ciency. Then, these photoelectrons are
accelerated by an electric eld of a few kilovolts or more, applied to the photocathode; they are bom-
barded onto the target semiconductor where electron-hole pairs are generated in proportion to the
incident energy and number of the photoelectrons. This is called "Electron Bombardment Gain". A
typical relation between the electron bombardment gain and the photocathode supply voltage is plot-
ted in Figure 3.48. In principle, this electron bombardment gain is proportional to the photocathode
supply voltage. However, there is actually a loss of energy in the electron bombardment due to the
insensitive surface layer of the semiconductor (see section 4.3), so the proportional relation does not
hold at a low voltage. In Figure 3.48, the voltage where the dotted line intersects the horizontal axis
is called the photocathode threshold voltage [Vth]. Electron bombardment gain increases in propor-
tion to the electron incident energy when the photocathode supply voltage is well beyond threshold.

113
fi
fi
fi
fi
ffi
fi
fi
fi

Figure 3.47 Schematic of a HPD detector. Figure 3.48 Electron bombardment gain as a func-
tion of the accelerating potential.

The electron bombardment gain of a HPD corresponds to the gain attained by the rst dynode of a
PMTs. A typical bombardment gain Gb is beyond 1000 up to 1500, which is much higher than the
multiplication factor of the rst dynode in conventional PMTs. In the next section, we will show how
this feature signi cantly reduces the gain uctuation of the detector, leading to a better performance
in terms of noise.
The cluster of secondary electrons resulting from electron bombardment is further multiplied by
the avalanche gain in the semiconductor (section 4.3). The gain Gd of the avalanche diode depends
on the supply voltage and has a typical value of 50-80. Then, the HPD total gain G is calculated as
follows:

G = Gb × Gd, (3.4.24)

and has a typical value of 5-10×104.


HPDs are superior to conventional photomultiplier tubes in many aspects thanks to the extremely
low uctuation in the total gain, because of the high value of Gb. Linearity and pulse height resolution
of HPDs are de nitely better than those of PMTs, yet HPDs have a lower gain and a limited sensitive
area.

NOISE IN PHOTOMULTIPLIER TUBES

The most part of the discussion concerning noise in vacuum photodiodes also holds for photomul-
tiplier tubes. The dark current emitted by the cathode is still given by the Richardson-Dushman
equation. Yet in case of PMTs the thermally emitted electrons undergo multiplication, as much as the
signal electrons do. Hence, the control of thermionic emission is especially important. In fact, even if
the average values of the dark current can be removed, its stochastic variation, that is its shot noise,
cannot and sums up to the signal current.
An additional noise source, not present in phototubes, stems from the multiplication process,
which has a stochastic variation. In fact, the gain G is not a deterministic quantity, but is a ected by a
variance v = σ2, where σ is the standard variation of the values of G measured in a sequence of exper-
iments. What is most important is the relative variance (RV) of G, which is the ratio of the variance
to the square of the mean:

σG2
RVG = . (3.4.25)
⟨G⟩2

114
fl



fi
fi

fi

fl
fi

ff
The gain G derives from a sequence of electron multiplications by dynodes, which are assumed to
be independent processes. Therefore, its relative variance can be estimated from the contributions
deriving from the single processes. We recall that the variance of a quantity that results from inde-
pendent contributions can be calculated summing the variances of the single contributes. Neverthe-
less, the electron multiplication by dynodes, although independent from each other, take place in a
sequence, being the output on the nth input on the n+1th. Therefore, the relative variances, but that
of the 1st dynode, must be rescaled for the number of incoming electrons. Hence:

RV2 RV3 RVn


RVG = RV1 + + +…+ , (3.4.26)
δ1 δ1 δ δ1 δ n−2

where δ1 is the electron multiplication factor of the rst dynode, which might be di erent from the
others, while δ is the multiplication factor of all the other dynodes. Note that the relative variance for
each dynode has been divided by the number of incoming electrons in order to refer the uctuations
of the bunch of electrons to the single electron that has triggered the cascade of multiplication pro-
cesses leading to G electrons at the anode.
Assuming that all the electron multiplication processes obey the Poisson probability distribution
(see Eq. 2.1.5 and 2.1.6) so that the variance is equal to the mean, from Eq. (3.4.25) we get:

1 1 1
RVG = + +…+ (3.4.27)
δ1 δ1 δ δ1 δ n−1

From Eq. (3.4.27) it is easy to understand that the rst dynode a ects the variation of the gain much
more than the others because the rst term in Eq. (3.4.27) is the largest one. Equation (3.4.27) sim-
pli es to:

δ1 ( δ ) δ1 ∑
1 1 1 1 n−1 1
RVG = 1 + + … + n−1 = (3.4.28)
δ k=0
δk

Using the properties of sums we get:

1 δ − δ 1−n 1 δn − 1
RVG = = (3.4.29)
δ1 δ − 1 δ1 (δ − 1) δ n−1

Considering the high value of δn, the term 1 in the numerator of the last equation can be neglected so
that we nally get:

1 δ
RVG ≈ , (3.4.30)
δ1 δ − 1

which provides a simple expression for the relative variance of the gain G. If all the dynodes have the
same multiplication factor δ, Eq (3.4.30) further simpli es to:

1
RVG = (3.4.31)
δ −1
Now we aim to calculate the signal to noise ratio at the cathode and at the anode to quantify the
consequences of the stochastic variability of the gain, measured by RVG. Let’s suppose that ⟨Ne⟩
electrons are emitted on average by the photocathode in a time interval Δt. Assuming a stationary
illumination, the relative variance of the number of photoelectrons at the cathode can be immediately
calculated:

115
fi
fi


fi



fi
fi
fi

ff

ff

fl

σN2e ⟨Ne⟩ 1
RVK = = = (3.4.32)
⟨Ne⟩ ⟨Ne⟩ ⟨Ne⟩
2 2

To calculate the variance of the number of electrons collected at the anode σA2 we need to sum the
variances of two processes: the emissions of photoelectrons and their ampli cation, which are as-
sumed to be independent of one another. To nd the rst contribute to the variance we start from the
standard deviation of primary electrons σNe , multiply it by the gain ⟨G⟩ , as if it were deterministic,
and take the square of the product, which actually gives the variance of the photoemission process
2
transferred to the anode. For the second term, we shall consider the variance of the gain σG weighted
(i.e multiplied) for the average number of electrons that undergo the multiplication process ⟨Ne⟩ ,
which is also assumed deterministic. Therefore, we get:

σA2 ≈ σN2e ⟨G⟩2 + ⟨Ne⟩ σG2 = ⟨Ne⟩ ⟨G⟩2 + ⟨Ne⟩ σG2 . (3.4.33)4

Since the average number of electrons collected at the anode is ⟨NA⟩ = ⟨Ne⟩ ⟨G⟩, its relative vari-
ance is:

σA2 σA2
RVA = = . (3.4.34)
⟨NA⟩ ⟨Ne⟩ ⟨G⟩
2 2
2

Substituting Eq. (3.4.32) into Eq. (3.4.33) yields:

1 1 σG2
RVA = + . (3.4.35)
⟨Ne⟩ ⟨Ne⟩ ⟨G⟩
2

Observing that the last factor in the previous equation is the relative variance of the gain RVG, the
value for RVA is then:

1
RVA = (1 + RVG), (3.4.36)
⟨Ne⟩

Finally, considering the relative variance at the cathode we get:

RVA = RVK (1 + RVG), (3.4.37)

which is consistent with the expectation that, if the gain were deterministic (RVG = 0), the relative
variance at the anode (RVA) would be the same of the one at the cathode (RVK). Yet, this is not the
case, due to the uctuation of the gain that introduces extra noise.
An important gure of merit of any ampli er, which accounts for the variance of the gain, is the so
called Noise Factor (F). The Noise Factor is the ratio between the [SNR]P at the input to the [SNR]P at
the output, measured in terms of power. The Noise Factor indicates how much the SNR will degrade
between input and output, and is de ned as follows:

⟨ e⟩ in-
4 You might wonder why in the second term of Eq. 3.4.33 the variance of the gain σG 2 is multiplied by N
2
stead of ⟨Ne⟩ . This happens because σG is the variance that applies to each electron produced by the cathode.
2

Since we have ⟨Ne⟩ electrons, in order to find the variance of the multiplication process, we need to sum ⟨Ne⟩
⟨ e⟩ G
2 , as it is required for independent processes. This corresponds to the value N σ 2 .
times the quantity σG

116

fi
fl






fi
fi
fi

fi

fi

(S /N )2in
F= (3.4.38)
(S /N )2out

Often the Noise Factor is expressed in logarithmic units through the parameter Noise Figure (NF):

NF = 10 log10(F ) = 10 log10(S /N )2in − 10 log10(S /N )2out [d B ] (3.4.38a)

A small Noise Factor means that the extra noise introduced by the ampli er does not compromise
the signal. To calculate the Noise Factor of a PMT we can take the ratio between the relative variance
at the anode (Eq. 3.4.37) and that at the cathode (Eq. 3.4.32):

1/RVK RVA
F= = = 1 + RVG, (3.4.39)
1/RVA RVK

Under the hypothesis that all the dynodes have the same multiplication factor, the Noise Factor be-
comes:

1 δ
F =1+ = . (3.4.40)
δ −1 δ −1
Nevertheless, from Eq. (3.4.27) we have already noted that the multiplication factor of the rst dyn-
ode is the one that mostly in uences the noise gure. For this reason, assuming that δ in Eq. (3.4.40)
refers to the rst dynode (which is the most important), it is evident that the noise gure can be im-
proved by increasing the multiplication factor. This explain why hybrid photodetectors show the best
performance in terms of gain stability.
Finally, to nd the r.m.s value of the shot noise current we can start from Eq. (3.4.11), that holds
at the photocathode:

[iShot ]K =
RMS
2eIsΔ f . (3.4.41)

Hence the signal to noise ratio is:

S Is Is
= = (3.4.41a)
N [iShot
RMS
]A 2eΔ f

The shot noise current at the anode must account for the gain G and its uctuation. This can be done
by applying to Eq. (3.4.41) the following modi cations: i) the electron charge e is replaced by G×e,
since a bunch of G electrons originates from any single photoelectron; ii) the signal current Is is re-
placed by Is×G considering the ampli cation; iii) the Noise Factor F is included into the equation.
This leads to:

[iShot ]A =
RMS
2eG 2 Is F Δ f (3.4.42)

The signal to noise ratio at the anode is then:

S G Is Is
= = . (3.4.43)
N [iShot
RMS
]A 2e F Δ f

117

fi
fi





fl


fi

fi
fi


fl
fi
fi

fi

From Eqs. (3.4.41a) and (3.4.43) it might seem that the multiplication process in PMTs only add ex-
tra noise. Yet, the great bene t of the gain can be understood by considering the thermal noise that
a ects PMTs, as well as vacuum photodiodes. In fact, thermal noise comes on stage after the signal
ampli cation has taken place, so that its adverse e ect is reduced by a factor G. To be more quanti-
tative, the r.m.s. noise current at the load resistor RL is:

( RL )
4k T
inRMS = 2eG 2 Is F + Δf (3.4.44)

The signal to noise ratio of the detection system is then:

S GI G 2 Is2
= RMSs = . (3.4.45)
(2eG Is F + RL )
N in 2 4kT
Δf

Putting numbers in Eq. (3.4.45) we can show that the gain G a ects the balance between the two
terms of the denominator in the square root. Yet, at variance respect to vacuum photodiodes, the
shot noise of the signal current cannot be neglected in PMTs, even in low light conditions when
PMTs are likely used.
The NEP can be nally calculated by replacing the signal current Is with the dark current Id in
Eq. (3.4.45). When we dealt with vacuum photodiodes we estimated that the shot noise of the dark
current is typically 8 orders of magnitude or more lower than the thermal noise. Hence, considering a
gain G = 104 - 106, which is typical for a photomultiplier, the thermal noise is comparable to the shot
noise of the ampli ed dark current in the denominator of Eq. (3.4.45). If the dark current is very low,
e.g. thanks to cooling, the thermal noise might still prevail. In that case the NEP is given by:

inRMS 1 hc 4k T
N EP = ≈ Δ f, (3.4.46)
Rλ G e ηλ RL

which is G times smaller that the NEP of a typical photodiode, making evident the advantage of using
PMTs in low light conditions and in photon counting systems. In other conditions, the shot noise of
the dark current might overcome the thermal noise. Then, Eq (3.4.46) does not hold, yet the NEP is
still orders of magnitude better that the one of a phototube without ampli cation.
Before concluding, it is worth noting that photomultipliers are far from being general purpose de-
tectors and are a ected by important hindrances: they are fragile, require high supply voltage and are
sensitive to magnetic elds, which might alter the trajectories of moving electrons. Further, when a
photomultiplier tube is operated in a pulse detection mode, spurious pulses with small amplitudes,
called afterpulses, may be observed after the signal pulses. Afterpulses mainly depends on the ionic
feedback, which a ects all vacuum tubes.

Applications of Photomultiplier Tubes


Photomultiplier tubes are widely used in precision analytical instruments, in many uorescence
spectroscopy systems, including DNA sequencers, ow cytometers, confocal and multiphoton micro-
scopes. In LIDAR systems they are used for their high sensitivity and fast response. PMTs are also
employed in many medical diagnostic devices.
Featuring the lowest jitter amongst PMTs, MCPs are the detector of choice, together with Single
Photon Avalanche Diodes (SPAD) in photon counting systems (see section 4.3). SPADs, being all
solid state detectors, are more robust and “quiet” (few dark counts) than phototubes, yet their sensi-

118
ff
fi
ff
fi
ff
fi

fi
fi



fl
ff

ff
fi

fl

tive area is de nitely lower than that of MCP-PMTs, which are still necessary in many applications
notwithstanding the high complexity and cost.
Last but not least, a large number of photomultiplier tubes are used in fundamental physics exper-
iments. As an example, the photograph at the beginning of the present chapter shows several thou-
sands of giant PMTs covering the walls, the oor and the ceiling of the Super-Kamiokande detector
(Kamioka Observatory, Institute for Cosmic Ray Research - Tokyo University). This huge installation
is made of a cylindrical stainless steel tank, 39.3m in diameter and 41.4m in height containing 50
ktons of water. It is the world's largest underground neutrino detector and has been designed to
study neutrino oscillations by measuring light from Cherenkov radiation.

STREAK CAMERA

The streak camera is the fastest available direct optical detector. In speci c conditions, it can re-
solve optical pulses with a time resolution down to about 100 fs (Full Width Half Maximum).
Figure 3.49 depicts the operation principle of a streak camera. The light pulse to be measured
passes through a slit, which might be the exit port of a spectrometer, and is imaged by an optical sys-
tem onto the photocathode of the streak tube, where photons are converted into electrons. Photo-
electrons undergo a sequence of processes: i) are accelerated by a pair of accelerating electrodes; ii)
pass through a pair of de ection plates; iii) are multiplied in a micro-channel plate (MCP); iv) hit the
phosphor screen of the streak tube; v) are reconverted into an optical image (streak image). In Figure
3.49 four optical pulses, which vary in terms of both time and space and have di erent intensities,
arrive at the photocathode passing through the slit. As the electrons pass between the sweep elec-
trodes, a high voltage ramp is applied between them at a timing synchronised to the incident light, as
shown in Figure 3.50. This initiates a high-speed sweep, which makes electrons leaving the photo-
cathode at earlier times arrive at the phosphor screen at a position close to the top of the phosphor

Figure 3.49 Schematic of a Streak Camera tube.

screen, while those electrons that leave the photocathode at later times arrive at a position closer to
the bottom of the screen (top⟺early electrons, bottom⟺late electrons). Hence, the time at which
the photoelectrons left the photocathode can be determined by their vertical position in the streak
image, while the horizontal position replicates the position on the slit, which might correspond to
the wavelength when a spectrometer is coupled to the streak camera. Finally, a low noise CCD cam-
era acquires the image of the phosphor screen, where the vertical direction serves as the time axis
and the brightness is proportional to the intensity of the incoming light.
Streak cameras are used for ultrafast photoluminescence spectroscopy, which is an important tool
for researches in material science, biology and photochemistry, just to mention the most important
applications. A typical wavelength-time map obtained from a streak camera coupled to a spectrome-
ter is shown in Figure 3.51. Integrating over the vertical axis (green rectangle in gure) leads to a

119
fi
fl
fl

fi
ff
fi

Figure 3.50 Operating principle of a Streak Camera.

Figure 3.51 Wavelength-time map at the output of a Streak


camera coupled to a spectrometer.

Figure 3.52 Spectra (a) and decay curves (b) of the photoluminescence emitted
by CdS and CdS semiconductors.

120
gated spectrum (e.g. the luminescence spectrum between 400 and 800 ps with reference to Figure
3.51), while integrating along the horizontal axis (blue rectangle) provides the decay dynamics with-
in a speci c spectral window (600-650 μm in gure).
Finally, Figure 3.52 shows the di erent spectral (a) and decay features (b) of the photolumines-
cence emitted by CdS and CdSe semiconductors, measured with a Streak camera. CdS and CdSe, be-
yond being employed as photodetectors (see next section) have been extensively used as pigments in
ne arts (from Monet, van Gogh and Picasso up to modern painters) and as colourants in glass lter
technology.

Photoconductive detectors
Photoconductive detectors were amongst the rst light detector to become commercially available
outside laboratories. For example they were largely used in photocells for system automation and in
light meters for photography from the 50s of the past century. This popularity was due to their ex-
treme simplicity and to the presence of an intrinsic ampli cation mechanism that allowed e cient
light detection and measurement even before the electronics era. Photoconductive detectors are made
by a thin slab of an intrinsic semiconductor deposited onto an insulating substrate and connected to
two electrodes (Figure 3.53). The absorption of photons induces the transition of electrons to the
conduction band, thus changing the resistivity of the semiconductor. This reduces the resistance of
the detector RPhc leading to a change in the current owing across the load resistance RL under the
e ect of the bias voltage VB in the circuit shown in Figure 3.54. A capacitor C can be used to remove
the CW voltage across RL, which might be rather high. Therefore, even if the primary signal is the
variation in the photoconductive resistance RPhc upon illumination, what is actually measured (Vout) is
the change of the voltage drop across the load resistance VR :
L

RL
VRL = VB . (3.4.47)
RL + RPhc
∂VRL RL
Vout = ΔRPhc = − VB ΔRPhc. (3.4.47a)
(RPhc + RL)
∂RPhc 2

Therefore, the output also depends on RL. The appropriate value of the load resistance RL that max-
imises the responsivity can be found by taking the derivative of Vout with respect to RL:

∂Vout RL − RPhc
= VB ΔRPhc, (3.4.48)
( Phc L)
∂RL R + R
3

The derivative is equal to zero for RL = RPhc. The condition of equal values of the resistances is said
“adapted load” and sets the working point of the detector. The output signal is then:

∂VRL VB ΔRPhc
Vout = ΔRPhc = − . (3.4.49)
∂RPhc 4 RPhc

Equation (3.4.49) shows that the output signal linearly increases with VB. Supply voltages up to
100 V are rather common to improve the responsivity and the NEP. Yet, a high supply voltage re-
quires a capacitor to remove the CW voltage from the output.
Photoconductive detectors are characterised by a threshold wavelength, like all quantum detectors.
The threshold wavelength depends on the band gap energy of the semiconductor, according to the
equation:

121
fi
ff
fi



ff

fi
fi

fl
fi

ffi
fi
hc
λ< . (3.4.50)
eEg

To calculate the responsivity we start considering the quantum e ciency, which depends on the
amount of light absorbed by the semiconductor. The fraction of absorbed radiation depends on the
thickness D (Figure 3.53)of the slab and on the Fresnel re ection coe cient r:

η = (1 − r) [1 − exp (−α D)], (3.4.51)

where α is the absorption coe cient of the semiconductor, whereas the Fresnel re ection is given by:

(n − 1)2
r = , (3.4.52)
(n + 1)2

being n the refraction index. It is worth noting that r could be rather high, considering the typical
values of n for common semiconductors (Table 3.5).

Figure 3.53 Slab of a semiconductor used as photo- Figure 3.54 Photoconductive detection circuit.
conductive detector.

Table 3.5

Semiconductor n r

CdS ≈ 2.5 18%

PbS, InSb ≈4 36%

HgCdTe ≈4 36%

Figure 3.55 shows the absorption coe cient of several semiconductor materials. Note the di er-
ence in slope between direct (CdS, GaAs, InGaAs) and indirect (Si, Ge) band gap semiconductors.
The rate of generation rg of couples e-h in the unit volume of a slab of semiconductor caused by an
irradiance E0 is given by:

ηE0 (W L) η E0 λ
rg = = . (3.4.53)
h ν (W L D) hc D

122



ffi

ffi

fl
ffi
ffi

fl

ff

Figure 3.55 Absorption coef cients of semiconductors for


infrared detectors.

The recombination rate in a semiconductor depends on many factor that includes the type of band
gap (direct or indirect) and the density of defects. Yet, when the carrier densities Δn and Δp in excess
to the stationary density of electrons and holes are small, recombination rate can be expressed by a
single parameter τc, which gives the average lifetime of the excess carriers.

rr = Δn /τc = Δp /τc. (3.4.54)

In equilibrium condition, we assume that the generation and recombination rates are equal to each
other (rg = rc). Hence, the excess carrier densities due to the absorption of photons are given by:

Δn = Δp = rg τc. (3.4.55)

The electrical conductivity in a semiconductor is given by:

σ = e(n μe + pμh ), (3.4.56)

where μe and μh are the mobility of electron and holes, respectively.


The change in conductivity upon light absorption is then given by:

Δσ = e(Δn μe + Δpμh) = erg τc(μe + μh), (3.4.57)

which is linearly dependent on the generation rate and, in turn, on the irradiance E0.
Finally, we can apply the Ohm’s law (i = G V) at the circuit of Figure 3.54, being Gphc the conduc-
tance of the detector and GL the one of the load resistor:

Gphc GL GL2
∂ Gphc ( Gphc + GL )
∂G ∂
Δi = ΔGphc ⋅ VB = ΔGphc ⋅ VB = ΔGphc ⋅ VB.
(Gphc + GL)
∂Gphc 2

(3.4.58)

123




fi

Then, assuming adapted load, that is Gphc = GL, we get:

( L ) 4
VB WD V WD V
Δi = ΔGphc ⋅ = Δσ B = e τc(μe + μh)rg B . (3.4.58a)
4 L 4

Introducing the expression of generation ratio given by Eq. (3.4.52), we get:

W η E0 λ VB
Δi = e τc(μe + μh) . (3.4.59)
L hc 4
It is worth noting that Eq (3.4.59) hides an ampli cation mechanism that can be easily revealed. In
fact, we can calculate the gain factor (Photoconductive gain) Mphc intrinsically present in this type of
detectors by taking the ratio of the electron ux (electrons per second) in the circuit to the number of
photons absorbed per unit time:

Δi
e
Mphc = . (3.4.59a)
rg (W DL)

From Eq. (3.4.58a) and (3.4.53) we get:

[ L 4 ] e [ (W L) η E0 λ ]
WD V 1 hc τ (μ + μ ) V
Mphc = e τc(μe + μh)rg B = c e 2 h B. (3.4.60)
L 4

The gain corresponds to the ratio between the number of electrons exiting the photoconductive
slab in a generic time interval Δt to the number of couples e-h generated inside in the same time. The
photoconductive gain Mphc is typically >> 1 and depends on the bias voltage VB. To make more clear
Eq. (3.4.60) we can assume that the mobilities of electrons and holes are equal to each other [µe = µh
= µ] and observe that the product of the electric eld in the semiconductor E =VB/(2 L) times the car-
rier mobility µe gives the drift velocity of carriers. In turn, the drift velocity divided by the length L
corresponds to the inverse of the drift time. Therefore the gain can be written as:

[ L 4 ]
μ V Eμ vdrif t τ
Mphc = 2 2 B τc = τc = τc = c . (3.4.61)
L L τdrif t

Hence, the photoconductive gain Mphc has an immediate interpretation: it is twice the ratio of the car-
rier lifetime τc to the time spent by carriers to cross the entire slab τdrift.
We might consider possible options to improve the gain. The bias voltage VB can be increased to
the limit of electrical breakdown. The length L of the slab can be reduced, but this decreases the ac-
tive area and forces a reduction in VB, not to overcome the breakdown electric eld. The carrier life-
time τc depends on trap state density. Trap states can capture either electrons or holes and make the
recombination time longer. Therefore, increasing the density of trap states (i.e. The density of de-
fects) improves the gain, but is detrimental for the time response of the detector, which typically
shows a time constant close to the recombination time τc.
The responsivity can be nally obtained from Equations (3.4.59) and (3.4.60):

Δi RL 1 ηλ VB eηλ
Rλ = = 2 e τc(μe + μh) RL = MPhc R. (3.4.62)
Pin L hc 4 hc L

The last expression explicitly shows the role of the photoconductive gain.

124

fi

fl
fi
fi

fi

The major noise sources in photoconductive detectors are the generation/recombination (g-r)
noise and the so called 1/f noise, which depends on the trapping and releasing of carriers. The power
spectral density S(f) of these noise sources is not white and their r.m.s. currents are given by:

4Δie MPhcΔ f
⟨ig−r ( f )2⟩ = , (3.4.63)
1 + 4π 2 f 2 τc2

B
⟨i1/f ( f )2⟩ = Δi Δf B ≅ 10−11. (3.4.64)
f

Figure 3.56 shows the schematic representation of the noise power spectral density S(f), which
includes the mentioned noise components at low and medium frequencies and the Johnson noise at
hight frequency.
Photoconductive detectors are largely employed to detect radiation in the infrared band up to
10-15 μm. The most common semiconductors used for the purpose are lead chalcogenides (PbS and
PbSe), for sensitivity up to 5 μm, the binary semiconductor InSb μm, up to 7 μm and the ternary
semiconductor HgCdTe up to 15 μm. In the case of HgCdTe, the band gap can be tuned by changing
the amount of Cadmium in the semiconductor alloy, thus controlling the spectral sensitivity. Due to
the low band gap of InSb and HgCdTe, cooling is often required to achieve high sensitivity with de-
tectors based on these materials. Figure 3.5 shows the Detectivity of photoconductive detectors
based on the most common semiconductor that are employed for the infrared spectral range.
The typical bandwidth of photoconductive detectors is in the range of few tens of KHz up to MHz,
even if detectors based on speci c semiconductors (e.g. InGaAs) can provide a much wider band-
width. Considering the e ect of the capacitor in the bias circuit and the e ect of the carrier lifetime τc,
the transfer function drops sharply at very low frequency, then it is almost at up to the cuto fre-
quency set by τc, as it is schematically shown in Figure 3.57.

Figure 3.56 Major noise contributes in different fre- Figure 3.57 Typical transfer function of a photocon-
quency bands. ductive detection system.

Junction photodetectors
Light detectors based on a pn junction (Figure 3.58) or other types of junctions are de nitely the
most performant and widespread detectors in the VIS-NIR spectral range. When a junction between
semiconductors with opposite doping is formed, the di usion of the majority carriers to the other
side of the junction leads to the creation of a depletion layer deprived of majority carriers across the
junction. The unbalanced charge of ionised acceptors (p-type) and donors (n-type) causes an electric

125

ff

fi

ff

ff
fl
fi

ff
Figure 3.58 np junction showing the charge deple-
tion region.

eld to establish. The electric eld induces a drift current that, at equilibrium, compensates the di u-
sion current of majority carriers that move toward the other side of the junction, so that no net cur-
rent occurs across the junction. The electric eld results in an electric potential (maximum to the n
side), represented in Figure 3.58 by the downward bending of the bands, from the p-type semicon-
ductor to the n-type. Finally, we note that equilibrium requires the same Fermi level in n-type and p-
type semiconductors.

PN PHOTODIODES

Let’s now assume that a low impedance conductive path is established at the terminal of the diode
and, consequently, across the pn junction. When a photon is absorbed within the depletion layer, the
generated electron and hole are separated and accelerated by the electric eld up to a drift velocity
vdrift, which is set by collisions with the ions of the lattice. Once the two carriers have reached the
ohmic region of the semiconductor, the charge transfer has completed. An electron is expelled from
the n-type semiconductor into the external circuit and, at the same time, an electron enters the p-
type semiconductor and recombines with a hole. Therefore, even if a couple of charges (e-h) was gen-
erated, a single electron charge ows in the external circuit per each absorbed photon. Conversely,
when a photon is absorbed in the ohmic region of the semiconductor, the couple (e-h) recombines,
resulting in no-response. If the photon is absorbed close to the junction, one of the two charge can
di use into the region where an electric eld is present. This results in a delayed response, as it will
discussed in the next subsection.
The schematic of a silicon photodiode is depicted in Figure 3.59, which shows the aluminum elec-
trodes and the pn junction made by a highly doped p type semiconductor (P+) di used within a low
doped n type bulk (N-). The N+ layer close to the aluminum metallisation provides an impedance
match and facilitates the passage of electrons from the semiconductor to the metal, thus improving
the performances of the device. The quantum e ciency depends on the re ectance at the air-silicon
interface, which can be mitigated by an antire ection coating, and by the thickness of the depletion
layer. Being silicon an indirect gap semiconductor, its absorption coe cient is rather low (α =
105-106), compared to the one of direct gap semiconductors like GaAs. This might limit the quantum
e ciency of pn junction photodiodes. Using a very simple model the quantum e ciency is:

η = (1 − r) ⋅ {exp(−α wp ) − exp[−α (wd + wp )]} ≈ (1 − r){1 − exp(−α wd )}, (3.4.65)

where r is the re ectance, α is the absorption coe cient, wp is the thickness of the p+ deposition,
which is usually very thin, and wd is the thickness of the depletion layer.
The current generated by the absorption of light is:

126
fi
ffi
ff
fl

fi
fl

fi
fi
fl
ffi

ffi

ffi
fi
fl

ffi
ff

ff
Figure 3.59 Schematic of a np photodiode with re-
verse bias.

Pin η E0 A eλ
iλ = ηe = , (3.4.66)
hν hc
where E0 is the irradiance and A is the active area. The internal responsivity in terms of current is
therefore:

[W]
iλ e A
RλI = = ηλ ≈ 0.8 ⋅ 106 ηλ ≈ 0.5 {λ μ m} , (3.4.67)
Pin hc

Where it has been highlighted that the current responsivity has a very simple expression when the
wavelength is expressed in microns. Therefore, in the VIS wavelength range, the typical responsivity
of a photodiode is 0.2-0.3 A/W, while it can reach 1 A/W in NIR range and speci cally at the wave-
length of the maximum transparency of optical bres (1.55 μm).
The equivalent circuit of a photodiode is depicted in Figure 3.60, where the current generator ac-
counts for the conversion of photons into electrons, the diode represents the V-I characteristic of the
pn junction, the capacitor CJ is the capacitance of the junction, RS is the series resistance, typically
low enough to be neglected, RSh is the shunt resistance, which is also negligible, being very high. The
external circuit is made by the load resistance RL and by the bias voltage generator VB, whose values
set the di erent operating regimes of the detector, as discussed hereafter.
a) Photovoltaic mode (VB = 0, RL = ∞)
In this operation mode, the photodiode is attached to a high impedance ampli er, without any
bias voltage. Therefore, the current iλ generated by the photon ux must pass through the diode
shown in the equivalent circuit. Taking advantage of the mentioned approximations for the
equivalent circuit and using the I-V characteristic of a diode one gets:

[ ( kT ) ]
eVd
iλ = id with id = i0 exp −1 , (3.4.68)

where id and Vd are the current and the tension across the diode, respectively and i0 is the reverse
current due to drift of minority carriers in reverse bias. The output voltage, which is the voltage
signal across the diode can be calculated using Eqs. (3.4.66) and (3.4.68):

( h c i0 )
kT η E0 A e λ
Vλ = Vd = log +1 . (3.4.69)
e

127
ff



fi

fl

fi
fi

When iλ ≫ i0, the term -1 in Eq. (3.4.69) can be canceled making evident the logarithmic re-
sponse of the photodiode with respect to the incident power, given by Pin = E0 A.

Vλ ∝ log (α η λ Pin), (3.4.70)

where α is a constant.
The photovoltaic operating mode is seldom used. Yet it can be convenient when a huge dynamic
range is required and when the behaviour of the human eye has to be simulated. This is the case,
for example, when the outdoor light level must be measured in order to take an appropriate ac-
tion.
b) Photoconductive mode (VB = +VH, RL = 50 Ω)
In photoconductive mode (Figure 3.61) a reverse bias is set across the diode and the current iλ
passes through the load resistance RL producing a voltage signal Vλ:

( )
η E0 A e λ η E0 A e λ
Vλ = i0 + RL ≈ RL. (3.4.71)
hc hc

The voltage responsivity RλV is linear versus the input power, provided that the reverse current i0
can be neglected, which is the case in normal conditions. Nevertheless, in precision radiometry, a
better linearity can be achieved by removing the bias voltage as discussed below.
c) Photoconductive mode (VB = 0, RL ≤ 50 Ω)

Figure 3.60 Equivalent circuit of a pn photodiode.

When the cathode of the diode is at ground potential and the anode is kept at zero potential
through a virtual ground, the diode current id vanishes, according to Eq. (3.4.68) with Vd  =  0,
and the only current passing through the load resistor is iλ. The choice of no bias can extend the
linearity toward very low light, when the reverse current of the diode (dark current) is compara-
ble with the signal current. However, this approach worsens the responsivity and the frequency
response of the detector, which indeed, in radiometry, is seldom a concern since quasi CW mea-
surements are usually done.
The typical readout circuit of a photodiode relies on a transimpedance ampli er shown in Figure
3.61, which converts the photocurrent into the output voltage at low impedance.

128

fi

The width of the depletion region can be somehow controlled with the bias voltage and is mod-
elled by a simple approach based on the equations of the electrostatics. More speci cally, we apply
the depletion approximation for the pn junction, which assumes that the charge densities in the p
and n regions close to the junction are step functions, whose amplitude is given by the density of ac-
ceptors and donors, respectively. Boundary conditions requires that the electric eld vanishes at the
edges of the depletion regions and is continuous at the junction, but its derivative. The electric po-
tential is set by the external bias VB, which is supposed to be much larger than the built-in potential
barrier Vbi of the diode. The built-in potential is a consequence of the electric eld across the junction
that generates the drift current equal to the di usion current at equilibrium:

( ni2 )
kT Nd Na
Vbi = log ≈ 0.7V (3.4.71a)
e

The photodiode is assumed to be degenerate along two directions, in such a way that it can be de-
scribed by a mono-dimensional model.
A double integration of the Gauss law in each region of the depletion layer gives:

d Ex (x) ρ(x)
= (3.4.72)
dx ε0 εr
ρ(x)
Ex (x) = x (3.4.73)
ε0 εr

ρ(x) x 2

V (x) = − Ex (x) d x = − , (3.4.74)
ε0 εr 2

where ρ is the charge density, given by −eNa or +eNd in the p and n type semiconductors, respective-
ly, Ex(x) is the x component of the electric eld and V(x) is the electric potential. Note that to calcu-
late the appropriate functions representing the electric eld and the potential, which are shown in
Figure 3.62, appropriate boundary conditions must be applied. However, the From Eq. (3.4.73) it is
evident that the slope of the electric eld is positive in the n region and negative in the p region
(Figure 3.62). Furthermore, the voltage drop across the n-type and p-type semiconductors can be
obtained from Eq. (3.4.74):

2
e Nd wn2 e Na wp
ΔVn = ΔVp = . (3.4.75)
ε0 εr 2 ε0 εr 2

Further, we impose charge balance and require that the Kirchho 's voltage law is satis ed, namely
that the voltage drop equals the external bias:

Na wp = Nd wn VB = ΔVp + ΔVn. (3.4.76)

Solving the system of Eqs. (3.4.75) and (3.4.76), we get the width of the depletion region in the
two semiconductor types:

Na 2 ε0 εr VB Nd 2 ε0 εr VB
wn = wp = . (3.4.77)
Na + Nd e Nd Na + Nd e Na

The total width of the depletion region is:

129


fi

fi

ff
fi

ff

fi

fi
fi
fi

2VB ϵ0 εr (Na + Nd)


wd = wn + wp = . (3.4.78)
e Nd Na

Equations (3.4.77) can be simpli ed, assuming that the doping of the p-type semiconductor is much
greater than that of the n-type, i.e. Na >> Nd, as it is the case in typical devices:

Figure 3.61 Bias circuit of a photodiode operated Figure 3.62 Charge density, electric eld and elec-
in photoconductive mode. tric potential in a pn junction derived using the
depletion approximation.

2 ε0 εr VB 2 ε0 εr Nd VB
wn = wp = . (3.4.79)
e Nd e N a2

It can be noted that the depletion region mainly extends in the n-type semiconductor, as it is
shown in Figure 3.62, which depicts the charge density, the electric eld and the potential drop in a
reverse biased diode.
The thickness of the depletion layer, which in uences the sensitivity of the detector, can be aug-
mented by increasing the reverse bias voltage, but this approach is not very e ective because wn and
wp, according to a rather optimistic model, depend on the square root of ΔV, which in turn is limited
by the occurrence of the junction breakdown. Moreover, the dependance of the width of the active
layer on the bias voltage might a ect the linearity of the response at high current for the change in
the voltage drop across the junction due to a mechanism similar to the one discussed in section 4.1
for the photomultipliers.
The temporal response of pn photodiodes is severely a ected by a phenomenon called “di usion
tail” that can be easily observed in response to short light pulses. In fact, in a simple pn photodiode a

130

fi
fi
ff


fl
ff
fi
ff

ff
Figure 3.63 Diffusion tail in the pulse response of a
photodiode.

fraction of the light is absorbed in the ohmic region of the semiconductor, where no electric eld is
present. When the generation of couples e-h takes place close to the junction, few carriers might dif-
fuse up to the depletion layer, where they are accelerated and collected at the terminals, contributing
to the response. Yet di usion is a slow process compared to drift (see section 2.3), and its character-
istic time is described by the equation:

d2
τdif f = . (3.4.80)
2D
τdi is quadratic in the di usion length d and inversely proportional to the di usion coe cient D. As a
matter of fact, due to this e ect, the current generated by a short light pulse is composed by a prompt
and a delayed response, as it is schematically depicted in Figure 3.63. The late arrival of carriers
worsens the frequency response and distorts the shape of light pulses.
To widen the active layer, while keeping a low bias voltage an enhanced photodiode con guration,
called PIN, has been devised ( Figure 3.64). In PIN photodiodes a thick layer of intrinsic (or very low
doped) semiconductor is sandwiched between the p and n type semiconductors. The intrinsic region

Figure 3.64 Schematic of a PIN photodiode. The pro le of


the electric eld is also shown.

131
fi

ff
ff
ff

fi

ff
ffi
fi

fi
is completely depleted with just a voltage drop of few volts and the size of the active layer is almost
independent on the bias voltage. Quantum e ciency is therefore:

η = (1 − r) ⋅ {exp(−α wp ) − exp[−α (wi + wp )]} ≈ (1 − r){1 − exp(−α wi )}, (3.4.81)

where wi is the width of the intrinsic layer. In PIN photodiode quantum e ciency can reach 80% in
the NIR spectral region; further, their time response is much less degraded by the di usion tail, com-
pared to pn photodiodes.
The thickness of the intrinsic layer depends on the semiconductor type. In silicon PIN diode, wi can
be up to 20 μm. In photodiodes made of direct band gap semiconductors wi can be an order of mag-
nitude lower (2 μm) because of the much higher absorption coe cient. Figure 3.65 shows the ab-
sorption coe cient as a function of the wavelength for common semiconductors used for photonics
devices.
Amongst the semiconductors shown in Figure 3.65, InxGa1-xAs is the most interesting for ultra-
fast telecom photodetectors. Typically, the lattice parameter x is set to 0.53 in order to obtain lattice
matching. In this case, Eg = 0.74 eV and λth = 1.68 μm, which matches the telecom window (1.5-1.6
μm). In heterojunction photodiodes, made with intrinsic InGaAs sandwiched between two layers of
doped InP (Figure 3.66), the charge generation is con ned to the InGaAs intrinsic layer, since the
energy gap of InP is Eg = 1.34 eV such as λth = 0.92 μm, which is outside the telecom window.

Figure 3.65 Absorption coef cient of different semiconductors used for


junction detectors.

The heterojunction reduces the dark current and avoids the di usion tail, which is virtually absent.
The bandwidth of a PIN photodiode depends on two main factors: i) the time τdrift required by pho-
ton generated carriers to travel across the depletion layer; ii) the RC time constant τRC resulting from
the junction capacitance and the load resistance:

wi Aϵ0 ϵr
τdrif t = τRC = RL , (3.4.82)
Vdrif t wi

In Eq. (3.4.82) Vdrift, is the drift velocity, which typically reaches the saturation value Vsat = 105 m/s,
A is the area of the junction and ε0 εr are the vacuum and relative electric permittivities.
Being τdrift and τRC independent quantities, the e ective time constant τ can be obtained from the
equation:

132

ffi

fi

ffi

ff
fi
ff

ffi
ffi

ff

Figure 3.66 Heterojunction PIN photodiode.

2 2
τ = τdrif t + τRC . (3.4.83)

Since τdrift increases with wi, while the opposite occurs for τRC, the optimal thickness of the intrinsic
layer wi results from a trade-o between two competing mechanisms. As a consequence, τ is mini-
mum when the thickness wi makes τdrift and τRC equal to each other. Assuming a single pole transfer
function, the output voltage for a modulated signal, as a function of the frequency, is:

iλ RL
Vout ( f ) = . (3.4.84)
1 + 4π 2 f 2 τ 2

As an example, we can calculate the cuto frequency for a PIN diode made of InGaAs, which is the
typical semiconductor material for fast detectors, featuring a bandwidth of several tens of GHz. As-
suming:
wi = 2 μm Vdrift = 105 m/s A = 100 × 100 μm2 εr = 12 RL = 50 Ω,
we get τdrift = 20 ps and τRC = 26 ps, which results in a time constant τ = 32 ps and a bandwidth of
4.8 GHz. Much better performances can be obtained with an optimised design and special con gura-
tions.
Finally, Table 3.4 lists the main characteristics of PIN photodiodes made of common semiconduc-
tors.

Table 3.4 Main characteristics of PIN Photodiodes - from Govind P.


Agrawal, Fiber-optic communication systems, 4th ed., Wiley, 2010.

133




ff

ff


fi
AVALANCHE PHOTODIODES

In low light conditions, the limit to the sensitivity (NEP) of junction photodetectors is set by the

Figure 3.67 Operating principle of avalanche pho- Figure 3.68 Plot of the impact ionization coef -
todiodes. cients as a function of the electric eld.

thermal noise of the load resistance. Therefore, in order to improve their sensitivity, internal ampli -
cation can be obtained by exploiting a charge multiplication mechanism, similar to the one discussed
for photomultipliers. The operating principle of avalanche photodiodes is shown in Figure 3.67.
When a high reverse bias is applied to a pn junction, primary electrons and holes generated within
the depletion layer can acquire enough kinetic energy from the electric eld to generate secondary
couples e-h by impact ionisation. A sequence of impact ionisation events can lead to an avalanche
process that results in a multiplication factor M up to 100 or more. This increases the responsivity
and reduces the NEP.
The process can be quantitatively described considering the impact ionisation coe cients for elec-
trons and holes that depend on the semiconductor type and on the electric eld. From (Figure 3.68)
it can be noted that the impact ionisation coe cients are di erent for electrons and holes.
In actual devices the generation and multiplication regions are separated, being intrinsic the rst
and p or n doped the second. The schematic of an avalanche photodiode is depicted in Figure 3.69,
which also shows the pro le of the electric eld.
The multiplication process can be modelled through a simple system of di erential equations that
describe the current of electrons and holes within the gain region:

d ie
= αeie + αhih
dx
(3.4.85)
di
− h = αeie + αhih
dx
where ie and ih are the electron and hole currents, respectively, while αe and αh are the electron and
hole impact ionisation coe cients. Charge conservation requires that the total current is a constant
at any x coordinate:

134

fi
ffi
fi
f
fi
ffi

ff

fi
ff
fi
ffi

fi
f
Figure 3.69 Schematic of an avalanche photodiode

I = ie(x) + ih(x) = const. (3.4.86)

Equation (3.4.86) can be used to recast the system of equations into a single linear di erential equa-
tion, in the unknown quantity ie(x):

d ie(x)
= (αe − αh ) ie(x) + αh I. (3.4.87)
dx
The solution of equation (3.4.87) requires a boundary condition, which can be easily found. With
reference to Figure 3.69, since multiplication occurs in a p type semiconductor, the hole current
must be zero at the right boundary of the multiplication layer, where the total current consists only of
electrons injected in the n-side of the junction:

ih(d ) = 0 ie(d ) = I. (3.4.88)

Using the condition (3.4.88) one gets:

exp[(αe − αh) x ]
αe − αh exp α − α d ( αe − αh )
αh I αh I
ie (x) = − + I+ . (3.4.89)
[ e
( h) ]

We can set αh/αe = kA and we can assume kA ≪ 1 when a p type semiconductor (e.g. GaAs) is used as
gain material. With this substitution, Eq. (3.4.89) simpli es to:

{ A }
I k − exp[αe (kA − 1) (d − x)]
ie (x) = . (3.4.90)
kA − 1

Finally, the multiplication factor M can be de ned as the ratio of the electron current exiting the gain
region to the electron current entering that region:

ie(d ) kA − 1
M= = . (3.4.91)
ie(0) kA − exp[(kA − 1) αe d ]

135




fi


fi

ff

We can explore the behaviour of M considering two limiting cases. If we assume that the multiplica-
tion of holes is negligible, according to the hypothesis kA ≪ 1, no positive feedback occurs from holes
traveling left and we get:

αh = 0 ⇒ kA = 0 ⇒ M = ex p(αe d ). (3.4.92)

In this case, electron multiplication is a rst order process described by the equation
die(x)
= αe ie(x), which gives an exponential growth of the current within the multiplication region.
dx
As a second case, we can assume that the electron and hole multiplication factors are equal to each
other. Equation (3.4.86) becomes an indeterminate form that can be evaluated using the L'Hôpital's
rule:

1
αh → αe ⇒ kA → 1 ⇒ M→ . (3.4.93)
(1 − αe d )

In this case, the multiplication factor diverges for d → 1/αe. This behaviour depends on the positive
feedback due to charge multiplication by holes which adds to that of electrons. This gives rise to a
resonant process, which is intrinsically unstable. A very high gain can be achieved, but noise increas-
es altogether. For this reason, semiconductors characterised by the conditions αe ≫ αh or αh ≫ αe
are used for the gain medium in practical devices.
The responsivity of an avalanche photodiode, as a function of the frequency f is given by:

η  e  λ
RAPD( f ) = M( f ) Rλ = M( f ) (3.4.94)
hc
The multiplication factor M(f) for a modulated light signal depends on the CW multiplication fac-
tor M0 and on the e ective carrier transit time τe:

M0
M( f ) = . (3.4.95)
(1 + 4π 2 f 2 τe2 M02 )1/2

When multiplication is mainly produced by electrons, namely kA ≪ 1, the e ective transit time is:

τe = CA kA τdrif t with CA ≅ 1. (3.4.96)

The resulting bandwidth, assuming a single pole transfer function, is then:

1 1
Δf = ≅ . (3.4.97)
2π τe M0 2π τdrif t kA M0

The bandwidth improves when kA ≪ 1 , con rming that preferred semiconductors for the gain
medium are those with strongly asymmetric ionisation coe cients. Eq. (3.4.97) shows the trade-o
between the APD gain M0 and the bandwidth ∆f, namely the unavoidable balance between speed and
sensitivity. Yet, the factor kA ≪ 1 reduces the penalty of the multiplication factor M0 in the denomi-
nator of eq. (3.4.97), allowing APDs to reach a bandwidth comparable to the one of PIN photodiodes.
Table 3.5 gives typical characteristics of avalanche photodiodes made of di erent semiconductors.
The most interesting case is the one of InGaAs APDs. Since InGaAs is a rather low band gap material,
it can hardly sustain the high electric eld required for an e ective charge multiplication. Therefore
InGaAs APDs rely on InP for multiplication in heterostructure devices. In InP, αh > αe, hence multipli-

136

ff





fi
fi

fi


ffi
ff

ff
ff

Table 3.5 Main parameters of APDs based on different semiconduc-


tors - from Govind P. Agrawal, Fiber-optic communication systems,
4th ed., Wiley, 2010.

cation takes place is a n-type semiconductor inside a SAM (Separate Absorption and Multiplication)
APD, which is shown in Figure 3.70. The device on the right side of Figure 3.70 is a SAGM APD,
featuring an extra grading layer made of InGaAsP. This layer provides some matching between In-
GaAs and InP to facilitate the charge transfer between materials with di erent band gaps.

Figure 3.70 Design of (a) SAM and (b) SAGM APDs contain-
ing separate absorption, multiplication, and grading regions.

NOISE IN PHOTODIODES

In PIN photodiodes the r.m.s. noise current is given by the contributions of the shot noise and the
thermal noise, as it was seen for photoemissive detectors:

⟨in2⟩ = [2e (⟨is⟩ + ⟨idark ⟩) + 4kBT /RL ] Δ f , (3.4.98)

where is is the signal current while idark is the current due to thermally generated couples e-h within
the depletion region.
To give an estimate of the dark current we might consider that the charge generation rate per unit
volume of the depletion region is given by the equation:

ni
[s m ],
−1 −3
rg = (3.4.99)
2τc

137

ff

where τc is the minority carrier lifetime. Therefore, the overall generation rate can be obtained by tak-
ing the product of rg times the volume of the depletion region Vd:

Rg = rgVd (3.4.100)

In indirect gap semiconductors, the carrier lifetime τc mainly depends on the density of defects. In
silicon, τc can reach the value of seconds or more in the best quality material, however, a reasonable
lifetime for detector grade silicon is τc ≅ 10 ms. The intrinsic carrier density @ 300 K is ni = 1.45·1016
m-3. Assuming a thickness of the depletion layer w = 10 μm and a diameter of the junction D = 100
μm, a rather optimistic estimate of the dark current gives:

Idark ≈ 10−14 A. (3.4.101)

Therefore, the shot noise term due to the dark current within the square brackets in Eq. (3.4.98) is:

2 e Idark = 3 ⋅ 10−33 A 2 s, (3.4.102)

which must be compared with the thermal noise term. At room temperature, for a value of the load
resistance RL = 50 Ω, one gets:

4k T
≈ 3 ⋅ 10−22 A 2 s. (3.4.103)
RL

It can be observed that thermal noise is several order of magnitude higher than the shot noise, as it
was in the case of the vacuum photodiode. Therefore, in low light conditions, when also the shot
noise of the signal current can be neglected, the signal to noise ratio is:

isRMS RL RL ηeλ
SN R = = Rλ Pin = Pin , (3.4.104)
⟨in2⟩ 4 kBT Δ f 4k TΔ f h c

or equivalently, in terms of power:

average signal power ⟨is⟩2 R2 P2 R


(SN R )P = = 2 = λ in L , (3.4.105)
noise power ⟨in ⟩ 4 kBT Δ f

where Rλ is the responsivity. The NEP* can be easily found by setting SNR=1 in Eq. (3.4.104) and
solving for Pin / Δ f:

Pin 1 4kBT hc 4kBT


N EP* = = = . (3.4.106)
Δf Rλ RL ηeλ RL

Hence, the detectivity is:

1 ηeλ RL
D* = = (3.4.107)
N EP hc 4kBT

To calculate the noise performances of avalanche photodiodes, we must take into account an addi-
tional noise source due to the variance of the multiplication factor M. The excess Noise Factor FA (see

138


section 4.1) can be found from an appropriate theory for multiplication noise. It turns out that FA
strongly depends on the the ratio kA and is given by the equation:

( M)
1
FA(M ) = kA M + (1 − kA) 2 − (3.4.108)

As already observed, the lowest noise can be obtained when kA ≈ 0. In this condition one gets:

1
FA(M ) = 2 − ≈2 if M ≫ 1. (3.4.109)
M
Figure 3.71 shows the excess Noise Factor FA for avalanche photodiodes, as a function of M, for dif-
ferent values of the ratio kA.
The Noise Figure (Eq. 3.4.38a) of an ampli er is often expressed in logarithmic units. The minimum
Noise Figure for an avalanche photodiode is then 3 dB.
The r.m.s. noise current in APDs can be obtained from Eq (3.4.98), taking into account the multi-
plication process and the Noise Factor. Each primary electron undergoing an avalanche multiplication
becomes a bunch of M electrons. Therefore the electron charge in Eq. (3.4.98) must be replaced with
e M. Then, we need to consider that both the signal and the dark currents are multiplied by the factor
M. Hence, the r.m.s. noise current is:

⟨in2⟩ = [2e M (⟨is⟩ + ⟨idark ⟩) FA + 4kBT /RL ] Δ f


2
(3.4.110)

The signal to noise ratio is given by:

isRMS isRMS MRλ Pin


SN R = = = ,
⟨in2⟩ ⟨ishot
2 ⟩ + ⟨iT2 ⟩ [ 2e M 2 (Rλ Pin + Idark)FA + (4kBT /RL )] Δ f
(3.4.111)

which simpli es to:

RL
SN R = M Rλ Pin , (3.4.112)
4 kBT Δ f

when the thermal noise prevails, which is always the case in low light conditions. Then the NEP* and
the Detectivity* are:

1 hc 4kBT
N EP* = , (3.4.113)
M ηeλ RL

1 ηeλ RL
D* = =M . (3.4.114)
N EP hc 4 kBT

Eq. (3.4.113) shows that the avalanche multiplication results in an improvement of the NEP* by a
factor M, with respect to a pn or a PIN photodiode. Also the SNR increases by a factor M due to the
ampli cation process. Nevertheless, one should always bear in mind that the previous result has been
achieved assuming a negligible shot noise of the signal current. If the signal is rather high, so that
this condition does not hold, ampli cation can worsen the SNR due to the excess Noise Factor FA.

139
fi

fi



fi


fi

Figure 3.71 Excess noise factor as a function of the ava-


lanche gain.

HOMODYNE DETECTION

The transmission of information through an optical link requires the modulation of a carrier using
di erent formats. For many years, since the beginning of the optical era, the transmission of digital
information through optical communication systems was based on the simplest amplitude modula-
tion scheme, called Pulse Code Modulation (PCM). In this scheme, a de ned time slot is allocated to
each bit of a stream and bits 0 and 1 are encoded by the absence or presence of an optical eld within
the time slot. This modulation scheme, which is also called amplitude-shift keying (ASK), is very
simple, but its e ciency in terms of occupation of the available bandwidth is rather poor. Figure
3.72 shows the most common modulation formats for PCM transmission. In the rst format, which
is called Non-Return-to-Zero (NRZ), bits 1 occupy the whole slot, while in the Return-to-Zero (RZ)
format the optical signal occupies only half of the slot for 1 bits. RZ format allows the receiver to
maintain the synchronism with the bit stream, but requires a larger bandwidth respect to NRZ. In
Manchester format, half of the slot is occupied by the optical signal also for 0 bits, making easier the
extraction of the clock from the bit stream.
Modern optical systems use complex modulation formats that rely on the interplay of all the prop-
erties of the electromagnetic eld: amplitude, phase and polarisation. An extensive discussion of the
modulation formats in optical communication is far beyond the scope of the present paragraph. Nev-
ertheless, it is worth considering brie y the concept of coherent transmission systems, in which the
information is encoded into the phase of an optical eld. The corresponding modulation format is
called phase-shift keying (PSK).
Given an optical carrier of angular frequency ωc that transmit information along an optical channel,
the electric eld at the receiver is represented by the equation:

ec(t) = Ec cos [ωc t + θ (t)], (3.4.115)

where Ec is the amplitude of the carrier and θ(t) represents the phase modulation that encodes the
information to be transmitted. To recover the information at the receiving unit, a phase sensitive de-
tector is required. Nevertheless, all the detectors seen so far are sensitive to the light intensity, which
means the square amplitude of the electric eld. Hence, in coherent optical system a di erent detec-
tion approach is required, which is based on the principle of interference. To this purpose, the electric

140
ff
fi
ffi

fi

fl

fi
fi

fi
fi
ff
fi

eld of a local oscillator having the angular frequency perfectly matched to that of the carrier is added
to the electric eld of the signal using an optical mixer, which could be, in the simplest implementa-
tion, a semitransparent mirror, as it is shown in Figure 3.73. This detection scheme is called Optical
Homodyne. The two elds interfere at the detector, which is typically either a PIN or an avalanche
photodiode. In gure Figure 3.73, two detectors collect complementary signals that can be used for
noise compensation. The coherent superposition of the electric elds is given by:

1
ecoher = {Ec cos [ωc t + θ (t)] + Elo cos [ωc t + φ]}, (3.4.116)
2

where 1/√2 accounts for the beamsplitter, Elo is the amplitude of the local oscillator and φ is a phase
constant that depends on the phase of the local oscillator and on the length of the optical paths. Be-
ing time independent, φ can be controlled to maximise the output signal, as it will become clear in
the following discussion.
The current intensity generated by each detector can be obtained by the product of the light power
times the responsivity Rd of the detector:

1
i(t) = ϵ0 c Ad Rd{Ec cos[ωc t + θ (t)] + Elo cos[ωc t + φ]}2 (3.4.117)
4
where Ad is the active area of the detector and the light irradiance has been calculated from the
square of the electric eld [see Eq. (1.1.4)].
Expanding all the terms in Eq. (3.4.117) and using the properties of trigonometric functions, one
gets:

1
i(t) = 8 ϵ0 c Ad Rd{Ec2 + Elo
2
+ 2Elo Ec cos[θ (t) − φ] + Ec2 cos[2ωc t + 2θ (t)] +
(3.4.118)
2
2Elo Ec cos[2ωc t + θ (t) + φ] + Elo cos[2ωc t + 2φ]}

Equation (3.4.118) contains two CW terms, which are proportional the the intensities of the local and
signal electric elds and three terms oscillating at an angular frequency twice that of the carrier. All

Figure 3.72 The most common PCM modulation Figure 3.73 Optical mixer used for homodyne de-
formats. tection.

these terms can be neglected. In fact, the CW terms can be easily removed with a lter, while the

141
fi

fi
fi
fi
fi
fi



fi

fi

terms oscillating at 2 ωc simply do not appear in the output, since no device can sustain a current at
such a high frequency (≈ 1014 Hz). Therefore, the only remaining term is:

1
i(t) = ϵ0 c Ad Rd Ec Elo cos[θ (t) − φ], (3.4.119)
4
or equivalently:

1
i(t) = R Pc Plo cos[θ (t) − φ]. (3.4.120)
2 d
Equation (3.4.120) demonstrates that the output current provides the required information, which
has been encoded in the phase θ(t) of the electric eld. Assuming that θ(t) oscillates around a con-
stant value θ0, which can depend on the lengths of optical paths and electrical cables, the phase can
be tuned such as [θ0 − φ = 0] in order to maximise the signal i(t).

Figure 3.74 Average optical power required to decode a bit


stream for different detector technologies.

It is worth noting that in homodyne detection an intrinsic ampli cation mechanism is embedded
in the multiplication of the amplitude of signal by the eld of the local oscillator. In fact, Elo is typical-
ly much higher that Ec, which might be very faint after a segment of optical bre in long-haul com-
munication links (hundreds of km). This large multiplication factor makes homodyne the most per-
formant detection scheme used in optical communication and whenever the electric eld of a signal
must be recovered instead of its intensity.
Figure 3.74 shows the average optical power that is required to decode a bit stream with an aver-
age Bit Error Rate (BER) equal to 10-9 as a function of the bit rate, for di erent technologies of the
receiver. The optical power is given in the logarithmic unit dBm: [1 dBm = 1 mW].

142

fi
fi
fi

ff
fi

fi

𝜑
In homodyne, the Fourier spectrum of the detected current is centred at zero frequency, which
might be a region a ected by severe noise. Therefore, in optical network often the local oscillator has
an angular frequency slightly di erent (hundred of MHz) from the one of the carrier. In this case, the
detection scheme is called Heterodyne and the center frequency of the signal is:

1
fHeterodyne = (ωlo − ωc). (3.4.121)

Except for this di erence, all the concepts and advantages shown for Homodyne also apply to Het-
erodyne, which is de nitely more common in practical applications.

143
ff
ff
fi

ff

Chapter 4

PIXELATED DETECTORS
J'ai trouvé le moyen de xer les images de la chambre obscure !
- Louis Daguerre

144

fi

Section 1

INTRODUCTION
In recent years a great progress has occurred in imaging detectors, leading to a huge increase in
performance and drop in cost. This has determined a pervasive presence of imaging sensors in pro-
fessional and consumer electronics. Further, the improvement in quality of pixelated sensors has
brought these devices to become amongst the most performant detectors routinely used in scienti c
research. In fact, pixelated detectors feature sensitivity close to the quantum limit, low noise, high
throughput and an intrinsic parallel architecture. Such a wealth of opportunities has paved the way to
the development of new methods in astronomy, microscopy and spectroscopy. This has resulted in a
widespread use of imaging detectors for research in fundamental physics, material science, biology
and medicine.
The devices presently available for scienti c imaging belong to two major families, namely Charge
Coupled Device (CCD) and Complementary Metal Oxide Semiconductor (CMOS). Even if great dif-
ferences exist between the two technologies, they share an important common feature. In fact, these
imaging detectors are based on the charge storage capability provided by the depletion layer of a
properly biased semiconductor substrate. The electrons generated by photons, proportionally to the
light intensity absorbed by silicon or other semiconductors, as InGaAs, are accumulated in Metal Ox-
ide Semiconductor (MOS) capacitors that form the pixels of the sensor. Then, after the exposure of
the detector to light, a charge transfer process takes place to bring the electrons accumulated in the
matrix of sensitive elements to a readout circuitry where charge-to-voltage conversion takes place.
The readout circuit provides an analogue signal that is bu ered, ampli ed and eventually converted
to a digital output. While the charge storage process is similar in CCD and CMOS sensors, the read-
out process is di erent in the two device families. In fact, a unique readout element characterises
CCDs, while in CMOS sensors each pixel has its own charge-to-voltage conversion unit, and the sen-
sor often includes ampli ers.
In this chapter we will study the basic operating principles of CCD and CMOS imaging detectors
and we will focus our attention to scienti c grade devices. Actually, when using pixelated detectors in
spectroscopy systems or other experimental set-ups, the choice of the sensor should be made by con-
sidering several parameters that include: the ll factor (i.e. the active area), the maximum charge that
can be stored per pixel, the noise gure and the operating speed.

145

ff
fi

fi
fi
fi
fi
ff

fi
fi
Section 3

THE MOS CAPACITOR


The MOS (Metal Oxide Semiconductor) capacitor is a two-terminal device that is essentially the cen-
tral part of an MOS transistor without source or drain (Figure 4.1). Despite its simplicity, the MOS
Capacitor allows us to study the most important concepts underlying the operation of eld e ect de-
vices, such as those of charge accumulation, which is the base for imaging detectors, and inversion,
which allows transistors to operate. Further, the study of the MOS capacitance is an important tool to
understand the behaviour of the MOS structures. Therefore, the MOS Capacitor is largely used in the
development and monitoring of semiconductor processes.

Figure 4.1 Schematic of a MOS Capacitor.

The discussion of the MOS Capacitor shall conveniently start from a band diagram that, in semi-
conductor physics, is the most powerful tool to understand the charge density and to visualise the
electrostatic potential. First of all we draw the energy band structure of each of the materials that
make up the MOS capacitor. Figure 4.2 shows the band diagrams of metal, insulator, and semicon-
ductor, respectively. Energy levels are referred to the vacuum level, which is the energy of an electron
at rest far away from our structure.
At absolute zero, the metal has lled states below the Fermi level Fm and empty states just above.
As we have already seen in section 4.1, the work function Φm = Evacuum - Fm is a fundamental property
of the metal, which represents the minimum energy required to promote an electron from the metal
into the vacuum, as already seen.
Insulators and semiconductors are materials that at zero temperature have a completely lled va-
lence band and a completely empty conduction band. The distinction between insulators and semi-
conductors is not a sharp one. Materials regarded as semiconductors tend to have a small bandgap,
typically close to 1-2 eV, and can be readily doped to make n- and p-type materials. Conversely, insu-
lators have a bandgap of several electron volts. Both semiconductors and insulators are characterised
by the value of the energy gap EG = EC - EV and the electron a nity χ. We have already seen (Section
3.4.1) that the electron a nity is de ned as the energy required to excite an electron from the con-
duction band to the vacuum level. Thus χ = Evacuum - EC. The value of χ depends upon the composition

146
ffi
fi

fi

ffi
fi
fi
ff
Figure 4.2 Band diagram of the MOS materials

and crystal lattice of the material. The electron a nity is almost una ected by doping and can be re-
garded as a property of the material, except in speci c cases in which surface e ects cannot be ne-
glected, as we have seen for NEA photocathodes. Also for a semiconductor, we can de ne the work
function Φs = Evacuum – EF, where EF is its Fermi level. However, EF, being within an energy forbidden
region, does not represent the energy required to extract electrons from the semiconductor, as it is
the case for metals. Further, EF is not a property of the semiconductor since it moves within the band
gap as a function of temperature and, above all, depends on doping. Taking advantage of this depen-
dence, EF will be used as a reference energy to calculate the local carrier density in a doped semicon-
ductor in equilibrium or subjected to a potential di erence and to an electric eld.

The ideal MOS capacitor


In the following discussion of the MOS capacitor, the semiconductor will be a single-crystal of sili-
con. At room temperature, we will assume EG(Si) = 1.12 eV and χS = 4.05 eV. Finally, the energy of
the Fermi level for intrinsic silicon can be located to a good approximation in the middle of the
bandgap: Ei =  EC -  EG/2. This result can be obtained assuming that the e ective mass of holes and
electrons are equal.
Even if several insulators are available in the semiconductor industry, MOS devices are almost in-
variably fabricated using thermal silicon dioxide (SiO2) as an insulator for its many advantages. These
include a simple, controllable growth process achieved by heating the wafer in oxygen; a nearly ideal
interface with silicon; and the ability to act as a dopant di usion barrier. In addition, the Si-Si02 inter-
face is likely the most intensively researched and best understood of any in semiconductor technolo-
gy. Despite the excellent properties of the Si-Si02 interface, a signi cant shortcoming of Si02 as a gate
insulator is the relatively low value of its dielectric constant (εSi = 3.9), which limits the charge stor-
age of the MOS capacitor, as will be seen in section 4.4.2.
In order to construct the band diagram of the MOS structure, we need to know how the bands are
aligned at the interfaces. Essentially, we are forming two heterojunctions and we must preserve the
quantities that are characteristics of the materials, allowing an energy shift of the band edges in the
di erent materials, to account for any charge transfer between them, which results in electric eld
and potential drop.
Assuming that the the vacuum level is constant, the electron a nity of the semiconductor and of
the insulator (properties of the materials) sets with good accuracy the position of the respective con-
duction bands. Hence, in our model of the MOS capacitor we impose that the relationship between
the conduction bands is unchanged when the interface is formed. That is, at the insulator-oxide inter-
face, we keep constant:

147
ff

ffi
ff
fi

ff
ffi
fi
ff
ff
fi
ff

fi
fi
EC (SiO2 ) − EC (Si ) = ΔEC = χs − χi (4.3.1)

At the metal-insulator interface, we assume that the relationship between the conduction band of
the insulator and the Fermi level of the metal is preserved when the junction is created; thus we
have:

EC (SiO2 ) − Fm = Φm − χi = Φ′m (4.3.2)

where Φ′m is known as the modi ed work function.


For the following discussion, we assume that the semiconductor bulk is at ground potential (refer-
ence), while the metal (gate) is kept to a controlled potential (VG) through a voltage generator. To
make things simple, we set VG = 0.
We start our discussion considering an ideal MOS capacitor, that is a capacitor made of materials
with Φs = Φm. We also assume that there are no surface charges at the semiconductor-insulator inter-
face and no bulk charges within the insulator.
Figure 4.3 shows the resulting band diagram for VG  =  0. To obtain this diagram, let us start by
imposing that the semiconductor bands are at (i.e. that there is no band bending in the semiconduc-
tor). We will provide a justi cation for this assumption thereafter. From the fundaments of electro-
statics [E = − ∇V], which states that the electric eld is proportional to the slope of the bands, the
assumption of at bands implies that no electric eld is present in the semiconductor: Es = 0. Fur-
ther, since no volume charge ρ is present in an ideal MOS capacitor, the Gauss theorem at the inter-
face semiconductor-insulator, tells us that no electric eld is present in the insulator as well. In fact:

ϵiℰi = ϵsℰs, (4.3.3)

where εi and εs are the dielectric constants of the insulator and the semiconductor, respectively. Thus
the bands are indeed at all over the device since the metal cannot sustain any electric eld.
In general, when a real MOS structure is created, the Fermi levels of metal and semiconductor
might di er from each other, leading to a non-equilibrium condition. The zero bias condition
previously assumed (VG = 0) corresponds to the presence of a low resistance path (short circuit) be-
tween the metal and the semiconductor. Therefore, notwithstanding the presence of the insulator, the
energy unbalance of electrons in the metal and the semiconductor would lead to an electron ux that
would quickly restore the equilibrium, leading to the accumulation of a net charge of opposite sign
within the metal and the semiconductor. As it happens in a normal capacitor, this would result in a
voltage drop across the insulator, leading to a slope in its conduction and valence bands. Therefore,

Figure 4.3 Band diagram of an ideal MOS capacitor.

148


ff

fl
fl
fi
fi


fl

fi

fi
fi
fi

fl

the at band condition would be lost. Such a self-bias would make more complex the discussion of
the di erent operating regimes of the MOS capacitor. Nevertheless, if we apply an appropriate poten-
tial to the gate, which is just the di erence in Fermi levels of the metal and the semiconductor, we
can prevent any charge transfer to occur, restoring the atband condition, as it is shown in Figure 4.4.
Suppose that Φm > Φs. Thus, the atband condition can be achieved with a positive bias of the gate:
1
VG = VFB >0. In fact, the electromotive force of the generator must contrast the tendency of electrons
to move from the semiconductor toward the metal, where they would reach a lower energy status,
due to the di erence in Fermi levels. Hence, we get:

qVG = qVFB = Φm1 − Φs = qϕms, (4.3.4)

where ϕms (volts) is known as the work function di erence.

Figure 4.4 Flat band voltage (VFB) applied to a MOS capaci-


tor to restore the at band condition.

For simplicity, from now on we will assume that the ideal MOS capacitor is obtained with VFB = 0.
This implies a precise choice of the metal and the doping of the semiconductor. If this is not the case,
and VFB  ≠  0, we will implicitly add VFB to the voltage VG applied to the gate, without any speci c
mention.
Starting from the ideal MOS capacitor ( at bands with VFB = 0) we will explore the behaviour of
the device for di erent gate bias voltages, either negative or positive. The semiconductor bulk is al-
ways grounded.

The accumulation regime


When we apply a negative voltage to the the gate (VG < 0), the metal Fermi level Fm is pulled up by
| qVG | . This requires a voltage drop across the system. Yet, since the metal cannot sustain any elec-
tric eld, this must establish in either the semiconductor, the insulator, or both. Actually, according
to Eq. (4.3.3), a nonzero eld must be present in both of the materials. Figure 4.5 shows the band
diagram in an ideal MOS capacitor (A) and in a MOS capacitor biased in the accumulation regime
(B). Figure 4.5 B) shows that most of the voltage establishes across the insulator, where the slope of
the bands is proportional to the electric eld, which is uniform as in a parallel plate capacitor, and
directed toward the metal. At the interface insulator semiconductor, the electric eld drops due to the

149
fl
fi
ff

ff

fl
ff
fi

fl
ff
fl
fi
ff

fl

fi

fi
Figure 4.5 MOS capacitor biased in accumulation regime
(B) compared to an ideal MOS capacitor (A).

ratio in the relative electric permittivities of the two materials (εs/εi ≈ 3), but does not vanish. The
holes in the p type semiconductor, being free carriers, are attracted by the electric eld toward the
interface with the insulator, where their density increases above the density in the bulk (pbulk = NA).
Hence, an excess of positive charge accumulates in the semiconductor within a layer of small, but not
negligible thickness. This brings the valence band closer to the Fermi level, which is assumed con-
stant all over the semiconductor. We will discuss this assumption later on. The actual hole density p
[m-3] can be obtained through equation:

( kT )
Ei − EF
p = ni exp (4.3.5)

The excess charge density Qp [C/cm2] equals the modulus of the negative charge density QG [C/
m2] on the metal and screens the bulk of the semiconductor from the surface eld.
This bias condition is known as accumulation, and leads to an excess hole density that adds to the
already large density of majority carriers. In the present chapter the accumulation regime will not be
considered any further because it is not relevant for the use of MOS capacitors as charge storage ele-
ments for image recording.

The depletion regime


When VG >0 the Fermi level of the metal is pulled down by qVG. The bands in the insulator and in
the semiconductor bend downward and the electric eld is now directed toward the bulk. The holes
in the semiconductor are pushed away from the interface, where their density becomes negligible
compared to the acceptor concentration. Hence, a depletion region is formed, that leaves a negative
charge, with volumetric density qNA, made by the ionised acceptors. Actually, using Eq. (4.3.5), it is
easy to demonstrate that, when the bands are bent more than a few kT, the hole density almost van-
ishes. The thickness W of the depletion layer increases with VG, and settles to the value required for
electric neutrality. That is:

QG = − Qd = qNaW with QG > 0 (4.3.6)

150


fi

fi

fi

where Qd [C/m2] is the the depletion charge density, which prevents the electric eld to extend further
in the bulk. This condition is known as depletion bias. Figure 4.6 A) shows the band diagram and
the charge density in a MOS capacitor in depletion regime. To write Eq. (4.3.6) we have implicitly
applied the depletion approximation, which was already assumed in Section 4.3 to discuss pn junc-
tion detectors.

Figure 4.6 MOS capacitor biased in depletion regime (A) and


at the threshold for strong inversion (B).

The inversion regime


As the gate voltage keeps increasing, not only does the depletion width grow, but the total band
bending in the semiconductor also increases. Thus, the conduction band is steadily brought closer to
the (constant) Fermi level, and the electron concentration increases as well, according to the equa-
tion:

( kT )
EF − Ei
n = ni exp (4.3.7)

The electron concentration is initially very small (after all, initially n = ni2 /Na ), but eventually it
becomes signi cant compared to that of the holes in the bulk (p = Na). When this is the case, we say
that we are at the threshold of strong inversion. Figure 4.6 B) shows the band diagram just at threshold.
Even in science, the de nition of a threshold may sometimes stem from a convention that sets a
sharp discrimination between two contiguous states with di erent features. We chose as the criterion
for a "signi cant" electron concentration the condition that the electron density n close to the inter-
face semiconductor-oxide is equal to the concentration of the holes p in the bulk. That is:

ninter face = pbulk = Na (4.3.8)

At the threshold for strong inversion (Figure 4.6 B) the charge balance must include the electron
charge. Hence Eq. (4.3.5) becomes:

QG = − Qd − Qn (4.3.9)

151

fi
fi




fi

ff
fi

where Qn [C/m2] < 0 is called inversion charge density because the gate voltage has created a layer of
electrons that has “inverted” the surface of the semiconductor. Actually, the predominant charge car-
rier has changed from holes to electrons at the interface. The gate voltage at the threshold of strong
inversion is given the special symbol VT.
Finally, Figure 4.7 A) shows the consequences of a further increase in gate bias above the thresh-
old voltage VT. The additional gate charge must be re ected in an increase in |Qd| and |Qn|. Howev-
er, an increase in |Qd| corresponds to a linear increase in the width of the depletion layer W, which
results in additional band (e.g. Ei) bending in the semiconductor. In turn, this causes a strong in-
crease in charge density |Qn|, because this quantity stems from the electron density at the interface,
which depends exponentially on EF - Ei, as stated by Eq. (4.3.7), here recalled for convenience:

[EF − Ei ]Inter face


ninter face = ni exp (4.3.10)
kT

Thus, any increment in QG beyond the threshold is re ected almost entirely by an increase in |Qn|,
while the depletion charge |Qd| hardly changes at all. The resulting distribution of the negative
charge in the semiconductor is depicted in Figure 4.7 A): you might note that the width of the deple-
tion layer is equal to the one at threshold, shown in Figure 4.6 B). Hence, the depletion charge is
bound to the threshold value QT = - q NA WT, while the electron density Qn at the interface undergoes
a large increase with VG.
Before proceeding with the mathematical analysis of the MOS capacitor, let us rst reexamine one
of the assumptions made in the preceding discussion. In constructing the band diagrams, we drew a
constant Fermi level EF in the semiconductor. Using EF rather than the quasi-Fermi levels Fn and Fp is
strictly correct only in equilibrium, yet, an external bias, either positive or negative, was applied to
the MOS capacitor. How can we accept this?
First of all, we need to consider that our analysis is done a long time after setting VG. This corre-
sponds to assume a stationary regime. Further, the current densities of electrons and holes in a semi-
conductor can be expressed in terms of the quasi-Fermi levels Fn and Fp by equations:

d Fn d Fp
J = Jn + Jp Jn = μn n Jp = μp n . (4.3.12)
dx dx
The presence of the insulator prevents any current from owing across the device. In fact, the resis-
tivity of SiO2 has been reported to be in the range 1012-1016 Ω·cm, leading to immeasurably small
currents for an MOS capacitor. Since Jn = Jp = 0, the quasi fermi levels Fn and Fp must be at (dFn/
dx = dFp/dx = 0). Therefore, Fn and Fp are equal to each other all over the semiconductor because they
coincide in the bulk, where no e ect of the bias can be perceived. In fact, the ohmic region of a doped
semiconductor cannot sustain any electric eld nor potential drop in absence of a current. Hence:

Fn = Fp = EF, (4.3.13)

When the current through the insulator is negligible and Eq. (4.3.13) applies, it is conventional to
speak of the MOS capacitor in equilibrium and electron and hole density are determined by a single
Fermi level EF.

152

ff


fi


fl
fl
fl
fi

fl

The deep depletion regime


Let us now discuss brie y the behaviour of the MOS capacitor in a non-equilibrium state. This can
be attained by sweeping VG rapidly from accumulation to VG > VT. In this case, the MOS capacitor
may be biased into deep depletion, as shown in Figure 4.7. B).
Let us suppose that deep depletion has been induced by sweeping rapidly VG at t = 0. Immediately
afterwards, the electron concentration at the interface is extremely small, consisting of only those
electrons that were in the p-region at t = 0 and were swept to the surface while the semiconductor
was being depleted. Thus, the electron quasi-Fermi level Fn is close to the valence band at the surface
and keeps at through the depletion region. Then, Fn will approach FP many di usion lengths deep in
the substrate. On the contrary, the quasi-Fermi level Fp remains at all over the semiconductor since
holes, being majority carriers, free to move, rearrange very quickly in order to reach the appropriate
density in all the regions of the semiconductor. Note that now we have: n p < ni2 in two distinct re-
gions: within the depletion region, and also deeper in the substrate where Fn < FP. Thus, net genera-
tion of electron-hole pairs will occur in both of these regions. Electrons generated in the depletion
region will be swept immediately to the interface, while those generated near the depletion region
edge may recombine or may di use to the edge of the depletion region and then be swept to the in-
terface. Both components contribute to raising the electron concentration toward the equilibrium.
Thus, the inversion charge will begin to build up and Fn will rise toward FP. Eventually, the MOS ca-
pacitor will relax to the equilibrium condition set by the gate bias VG, which corresponds to strong
inversion (Figure 4.7 A). In absence of light, the generation currents that ow though the semicon-
ductor may be very small in carefully processed MOS capacitors, thus relaxation may take minutes at
room temperature or even more at low temperature. In fact, the generation rate in silicon can be
strongly reduced by controlling the density of defects and con ning the most defective part, like the
interface with other materials, far from the depletion region, as we will see in section 4.5. The persis-
tence of the non-equilibrium condition for a long time is what makes the operation of devices such as
CCDs, CMOS sensors and DRAMs possible.

Figure 4.7 MOS capacitor in strong inversion (A) and deep


depletion (B).

153
fl
fl
ff

fi

fl
fl
ff
The capacitance of the MOS capacitor
After de ning the most important operating regimes for the MOS capacitor, we can explore the
behaviour of this device by measuring its capacitance with the simple apparatus depicted in Figure
4.8. The main voltage generator sets the static gate voltage VG, while the small signal voltage genera-
tor ṽs provides a low amplitude sinusoidal waveform, whose frequency can be switched from low
frequency (e.g. few hertz or less) to high frequency (kilohertz or more). The capacitance can be ob-
tained by measuring the small signal current ĩ and using the equation:


C= . (4.3.14)
2π f ṽs

Figure 4.9 shows a schematic plot of the capacitance as a function of the gate bias voltage VG. The
signal voltage ṽs modulates the free carriers in the metal (electrons) and in the semiconductor (elec-
trons and holes). When VG is negative, the MOS capacitor is in the accumulation regime Figure 4.10
A). This means that plenty of holes are available at the interface with the oxide. The system behaves
like a plane capacitor of thickness ti, lled with a dielectric of permittivity εi. Hence, its capacitance
per unit area C is given by:

ϵi
C= = Ci. (4.3.15)
ti

When the gate voltage VG reaches zero, the regime switches from accumulation to depletion Fig-
ure 4.10 B). This means that free carriers in the semiconductor have moved away from the interface
within a layer of thickness W. Now, the signal voltage modulates the electrons in the metal and the
holes at the border of the depletion layer in a thin layer dW. The capacitance decreases and can be
easily calculated considering the series of two capacitors of di erent thickness and electric permittivi-
ty: ti, εi for the oxide and W, εs for the semiconductor. Therefore, the capacitance in depletion regime
is:

1 1
C= = (4.3.16)
1/Ci + 1/Cdep ti /ϵi + W /ϵs

By increasing VG, the threshold for strong inversion is eventually reached. Then, we need to distin-
guish between low (LF) frequency and high (HF) frequency small signal stimulus. Let’s start from LF
modulation. The density of electrons at the interface with the oxide equals that of holes in the bulk.

Figure 4.8 Test circuit to measure the capacitance Figure 4.9 Capacitance as a function of the bias volt-
of a MOS capacitor. age VG measured with low and high frequency sig-
nals.

154

fi



fi
ff

Therefore, the signal voltage now modulates the electron density Qn (Figure 4.11 A). In fact, we have
seen that above threshold the depletion charge Qd does not change at all (Figure 4.10 C). Since the
modulated charge is now at the interface, like it was in the accumulation regime, the capacitance
quickly rises to the one of the insulator. However, since in p type semiconductor electrons are minori-
ty carriers, the described scenario requires that new electrons generate from thermal energy during
the positive half period of the signal waveform ṽs and recombine in the negative half period. Because
silicon is an indirect semiconductor, thermal generation and recombination are rather slow processes
requiring phonon exchange with the crystal lattice. In section 4.3 we have seen that the characteristic
time for these processes, i.e. the time constant τ, in detector grade silicon is milliseconds or more.
Hence, low frequency is required to observe the modulation of Qn. Conversely, in high frequency
regime (Figure 4.11 B), the capacitance of the MOS capacitor is clipped to the one reached at the
threshold value, since the modulated charge is still made by holes at the edge of the depletion layer,
whose thickness does not change with VG.

Figure 4.10 Capacitive behaviour of a MOS capacitor in different regimes: A) ac-


cumulation; B) depletion; C) strong inversion.

The discontinuity in the behaviour of the MOS capacitance at threshold, which indeed is not
abrupt like depicted in Figure 4.10, allows one to experimentally determine VT and to measure im-
portant parameters of the system.
Finally, we can consider a last measurement in which the gate voltage is switched very quickly
from zero to the nal value VG > VT and the analysis is made immediately afterward. In this case, not
enough time is given to Qn to build up and the capacitance continues to decrease with VG, as predict-
ed by Eq. (4.3.16), where W is now greater than WT. This last condition has been already called deep
depletion and will be extensively discussed in the next section. Deep depletion is a very important
non-equilibrium condition since image detectors requires this regime to accumulate photon generat-
ed electrons.

Figure 4.11 Charge modulation in a MOS capacitor above threshold at low


(A) and high (B) frequency.

155

fi

Section 4

APPROXIMATE ANALYSIS OF THE MOS CAPACI-


TOR
In this section we will derive simple equations that model the electrostatic behaviour of the ideal
MOS capacitor. Even if this discussion is a ected by several approximations, its simplicity makes
crystal clear the charge storage capability of this device and allows us to discuss the main features of
pixelated detectors like CCDs and CMOS sensors.
We will now develop quantitative expressions for important parameters such as VT and WT. We
will also obtain an approximate expression for the capacitance C, as a function of the bias voltage VG.
Finally, to determine the maximum charge that can be stored in the MOS capacitor we will write an
expression that binds the inversion charge to the gate voltage VG.
Several assumptions make the calculations very simple. First of all, the width of the depletion re-
gion W is assumed to increase with VG up to the threshold bias VT. Thereafter, W is kept constant to
the threshold value (W = WT). The charge density in the depletion region is modelled by a step func-
tion (depletion approximation), while the inversion charge density made by electrons at the interface
semiconductor-insulator is supposed to be be con ned within a layer of negligible thickness (delta
function). Finally, the device is treated as a 1D structure that can be modelled using the fundamental
equations of the electrostatics.

Equilibrium
We start by assuming that the MOS capacitor is biased in the depletion regime with a gate volt-
age VG lower than the threshold for strong inversion VT. The corresponding band diagram is the
one already shown in Figure 4.6 A), which is depicted for convenience in Figure 4.12 A), together
with the diagrams of the charge density ρ, the electric eld E and the electric potential ϕ. We can ob-
serve that the Fermi level EF is not close enough to the conduction band to yield a signi cant density
of electrons. Band diagrams represent the energy for electrons. Therefore, band bending downward
corresponds to a positive electric potential ϕ(x) inside the semiconductor. The origin of the x axis is
taken at the interface semiconductor-insulator and the electric potential is set to zero in the semicon-
ductor bulk, which is supposed to be grounded (see Figure 4.1 and Figure 4.4). We aim to write an
expression for the electric potential ϕ(x) in the semiconductor and ultimately to nd the dependance
of the potential at the interface ϕs = ϕ(0+) on the bias voltage VG. A pictorial view of ϕ(x) is shown in
Figure 4.12 A) with reference to the valence band.
The electric eld E can be obtained by integrating the Gauss law with appropriate boundary condi-
tions:

dℰx ρ qNa
∇⋅E = = =− . (4.4.1)
dx ϵs ϵs

Since the density of the depletion charge has been assumed uniform, the electric eld in the semi-
conductor is expressed by a linear function with negative slope, which is null in the bulk. Therefore,
integrating Eq(4.41) and setting Es(W) = 0 we get:

156

fi

ff

fi
fi

fi
fi

fi
ϵs ( W)
qNaW x
ℰs = 1− . (4.4.2)

The electric eld in the insulator can be obtained by taking into account the di erence in electric
permittivity at the interface semiconductor-insulator:

ϵs q NaW
ℰi = ℰs(0+) = , (4.4.3)
ϵi ϵi

The electric potential ϕs at the interface can be then calculated by integrating the electric eld, with
the above mentioned boundary condition [ϕ(W) = 0]:

qNaW W
ϕs = . (4.4.4)
ϵs 2

Thereafter, the potential continue to increase linearly into the insulator. You might note in Figure
4.12 D) that the slope of the potential changes at the interface because of the discontinuity of the
electric eld. The voltage drop within the insulator is given by:

q Na W
Vi = ℰi ti = ti. (4.4.5)
ϵi

The gate voltage must be given by the sum of the voltage drops in the the insulator and in the semi-
conductor:

VG = Vi + ϕs. (4.4.6)

Hence, the width of the depletion region W, as a function of the bias voltage VG, can be calculated by
substituting Eqs. (4.4.4) and (4.4.5) into equation (4.4.6):

q ϵi Na W 2 + 2 q ϵs ti Na W − 2 ϵi ϵs VG = 0. (4.4.7)

The positive solution is:

q ϵs ti Na − q 2 ϵs2 ti2 N a2 + 2 q ϵi2 ϵs Na VG


W=− . (4.4.8)
q ϵi Na

With simple algebra, Eq(4.4.8) can be simpli ed to:

ϵs 2VG Ci2
W= 1+ −1 , (4.4.9)
Ci q ϵs Na

where Ci = εi/ti is the capacitance per unit area of the insulator, as already seen [Eq. (4.3.13)].
The capacitance C of the MOS in depletion regime can be calculated from Eq. (4.3.16):

q ϵs Ci Na
C= . (4.4.10)
q ϵs Na (2 Ci2 VG + q ϵs Na)

157
fi

fi








fi

ff

fi

Equation (4.4.10) can be recast in a simpler form by de ning a new parameter V0 = qεs NA /Ci2 ,
which has the dimension of a voltage,

Ci
C= . (4.4.11)
1 + 2 VG /V0

The parameter V0 depends on both insulator capacitance (which in turn depends on insulator
thickness) and doping concentration. It determines how rapidly the capacitance decreases with gate
voltage; the smaller V0 is, the more rapid is this change. In other words, V0 sets the voltage scale for
capacitance change in the MOS system. The decrease in capacitance is faster in capacitors with low Na
and thin oxides.
Lets now nd an expression for the gate voltage that brings the capacitor into the inversion
regime. The threshold for strong inversion VT was previously de ned as the condition when the elec-
tron density at the interface semiconductor-insulator is equal to the hole density in the bulk. Hence,
to nd a quantitative expression for VT we need to set:

n[sem−ins] = (ni2 /Na) exp (qϕs /k T )


{n[sem−ins] = pbulk = Na
(4.4.12)

By equating the right hand side of the above equations and solving for ϕs, we get an expression for
the potential in x = 0 at the threshold for strong inversion:

N a2
( ni ) ( ni )
kT kT Na
ϕsinv = log =2 log . (4.4.13)
q 2 q

This quantity is given a special symbol ϕs , which means inversion potential. Using the expression:
inv

( 2 kT)
E
ni = NC NV exp − G , (4.4.14)

for the intrinsic carrier density ni, Eq. (4.4.13) becomes:

[ ( 2 k T )] ( NC NV )
kT Na EG kT Na EG
ϕsinv = 2 log exp =2 log + . (4.4.15)
q NC NV q q

Equations (4.4.13) and (4.4.15) show the linear dependence on temperature of the inversion potential,
which also depends on doping.
We can nd another signi cant expression for ϕsinv by expressing the energy di erence Ei - EF in
the bulk from the knowledge of the hole density (Na):

pbulk = Na = ni exp ( kT )
Ei − EF

(4.4.16)
⇒ Ei − EF = k T log ( na )
N
i

Comparing the second equation (4.4.16) with the expression (4.4.13) for the inversion potential ϕs
inv
it is straightforward to write:

158
fi
fi
fi


fi

fi
fi

ff

[Ei − EF ]bulk
ϕsinv = 2 . (4.4.17)
q

Equation (4.4.17) gives an immediate interpretation to the criterion for the threshold for strong in-
version: ϕs is the surface potential such that the Fermi level is as far above Ei at the interface as it is
inv
below Ei, in the bulk Figure 4.6 B).
Now, from Eq. (4.4.4) we can derive an expression for the width of the depletion layer at thresh-
old:

2 ϵs ϕsinv
WT = . (4.4.18)
q Na

Equation (4.4.18) allow us to estimate the threshold width: WT.


The required bias voltage VT is also easily found from Eq. (4.4.6):

q Na WT
VT = Vi + ϕsinv = ti + ϕsinv. (4.4.19)
ϵi

Using Eq. (4.4.18) we nally get:

1
VT = 2 q ϵs Na ϕsinv + ϕsinv, (4.4.20)
Ci

which expresses the threshold voltage VT as a function of known parameters of the MOS capacitor,
namely Ci and Na, and the temperature, which is embedded in the de nition of ϕs .
inv
The time has come to give some numbers for the main parameters that characterise a typical MOS
capacitor based on silicon, for which we assume a 50 nm SiO2 gate insulator and a p-type substrate
doped to 1022 m-3. Table 4.1 shows the expected parameters for such a device.

Table 4.1. Typical parameters of a silicon based MOS capacitor

Parameter Ci ϕs_inv WT VT V0

Value 6.91 × 10-4 F/m2 0.70 V 3.0 × 10-7 m 1.39 V 0.35 V

Non Equilibrium
We start considering a sudden sweep of the gate voltage to a positive value well above threshold.
The MOS capacitor is now in deep depletion, with a depletion layer larger than the one at equilibrium
(WT). This condition is unstable because at the semiconductor-insulator interface n p < ni2 . The ap-
propriate diagrams for charge density, electric eld and the potential are represented by continuous
lines in Figure 4.13.
While time passes, the e-h couples thermally generated in the depletion layer and in nearby region
provide the inversion charge Qn at the interface. This a ects the electric state of the capacitor and
brings it toward the strong inversion regime (equilibrium).
In the semiconductor, the electric eld and the potential are still given by Eqs. (4.4.3) and (4.4.4),
even if now W > WT. In order to calculate the electric eld in the insulator, we will treat the inversion
charge Qn as a delta function (the arrow in the diagram of ρ in Figure 4.13).

159

fi


fi


fi
fi
ff

fi

qNaW Qn
ℰi = − (4.4.25)
ϵi ϵ1

The depletion width is still related to the surface potential by:

2 ϵs ϕs
W= (4.4.26)
q NA

The gate voltage VG drops both on the insulator and on the semiconductor, as it was in equilibri-
um, but now the balance also includes the inversion charge Qn:

ti
VG = ϕs + Vi = ϕs + ℰi ti = (Q Na W − Qn) . (4.4.27)
ϵi

By substituting W from Eq. (4.4.26) into Eq. (4.4.27), one gets:

(qNa 2ϵs ϕs /qNa − Qn)


VG = ϕs + . (4.4.28)
Ci

Then, the terms q Na and 1/Ci can be brought under the square root:

Figure 4.12 Electrostatics for a MOS capacitor in Figure 4.13 Electrostatics for a MOS capacitor that
depletion regime. starts in deep depletion and relaxes in strong in-
version.

160




Qn
VG = ϕs + 2 (qϵs Na /Ci2) ϕs − . (4.4.29)
Ci

Finally, recalling the de nition of the parameter V0 = qεs NA /Ci2, one gets a more signi cant rela-
tionship from the applied gate voltage and the inversion charge Qn:

Qn
VG = ϕs + 2V0 ϕs − . (4.4.30)
Ci

The last equation is fundamental to describe the behaviour of a pixel of a CCD employed for image
acquisition. Solving Eq. (4.4.30) for the potential at the interface ϕs we can predict its decrease while
Qn increases:

V0 ( Ci )
Qn 2 Q
ϕs = VG + − V0 1+ VG + n − 1 with Qn < 0. (4.4.31)
Ci

The electrostatics of the system can be understood by considering the dotted lines in Figure 4.13.
The increase in Qn makes the depletion layer shrink. This, in turn, con nes the electric eld in a
smaller region (with the same slope), such as its value is now lower than before at the interface. Yet,
in the insulator the electric eld E has increased, due to the presence of Qn at the interface. The po-
tential ϕ also attains a lower value at the interface and reaches VG at the boundary with the metal
through a linear function with a high slope in the insulator. The increase in slope accounts for the
rise of the electric eld in the oxide. Finally, the charge balance is also altered when the equilibrium
is reached because QG must re ect the new value of the electric eld in the oxide (E=QG/σi). This
means that in strong inversion we have more negative charge in the semiconductor, with respect to
deep depletion, since |Qd + Qn| = QG.
Now, we can look at the system from a di erent point of view. The depletion layer can be regarded
as a potential well capable of storing electrons and we wonder how many of them can be accumulated
in a CCD pixel. The answer still come from Eq. (4.4.30), which can now be solved for Qn.

Qn = − Ci (VG − ϕs − 2 V0 ϕs ). (4.4.32)

The potential well can be represented with an inverted axis for ϕ, as depicted in Figure 4.14.
When the MOS capacitor is biased in deep depletion, the potential well has the maximum depth and
is empty. The newly generated inversion charge Qn can be modelled as a uid that start lling the
well. In this way, the rise of the level of uid in our pictorial view represents the “decrease” of the
potential. How much can our “tank” be lled? We might believe that the charge could reach the edge
of the well (corresponding to ϕs = 0), like the water in a bucket, but this is not true. In fact, we need
a residual potential ϕs to keep electrons separated from holes, which are getting closer and closer
while the width of the depletion layer reduces. What is the minimum potential required to prevent
recombination? We can assume that this potential is ϕsinv, which sets the maximum electron charge
that can be stored in the capacitor:

( )
Qn ma x = − Ci VG − ϕsinv − 2 V0 ϕsinv . (4.4.33)

The MOS system is now in strong inversion, lled up with the maximum charge compatible with
the voltage VG applied to the gate. In this condition, the semiconductor region close to the insulator

161

fi
fi

fi
fl

fi
ff
fl

fi

fi

fi
fl
fi

fi
fi
behaves like a forward biased “np” diode and electrons that continue to be generated either by ther-
mal energy or photons get lost in the bulk, where they recombine with holes. As a further argument
to model this “current” as the one in a diode, we can observe that in a silicon based MOS capacitor, a
typical value for ϕs is 0.7 V, similar to the voltage across a forward biased pn junction.
inv

Figure 4.14 Potential well provided by the depletion


region at the interface semiconductor-insulator.

162

Section 5

CHARGE COUPLED DEVICES


In previous sections we have seen that the MOS capacitor is an ideal component to store a packet
of charge. In this section we will learn how the charge can be transferred along an array of capacitors
that share the same semiconductor bulk. These two features are the main requirements for the opera-
tion of a CCD.
The CCD was invented by Willard Boyle and George E. Smith on the 8th of September 1969 at Bell
Labs in Murray Hill, NJ. The Bell Labs research facility, nowadays owned by Nokia, at those times
was managed by AT&T Co. the largest provider of telecom services in USA. Bell Labs hosted many of
the brightest scientists who have changed the technology scenario in the past century. Amongst
them, Clause Shannon, the father of the Information Theory, and several Nobel Laureates, like:
• Clinton Davisson (electron di raction - 1937)
• William Shockley (transistor - 1956)
• Arno Allan Penzias (cosmic background radiation - 1978)
• Steven Chu (laser cooling of atoms - 1997)
• Willard Boyle and George E. Smith (2009)
• Eric Betzig (super-resolution in PALM microscopy - 2014)
• Arthur Ashkin (optical tweezers - 2018)
• … and others.
If this does not seem enough to you, we can mention some of the inventions born at Bell Labs that
have revolutionised our life: the transistor (1947); the MOSFET (1960); the CO2 Laser (1964); the
CCD (1969); the C Language (1972) and the Quantum Cascade laser (1994).
The CCD was announced at the beginning of 1970 in a scienti c paper that already reported all of
its main features, clearly listed in the abstract:
“In this paper we describe a new semiconductor device concept. Basically, it consists of storing charge in potential
wells created at the surface of a semiconductor and moving the charge (representing information) over the surface by
moving the potential minima. We discuss schemes for creating, transferring, and detecting the presence or absence of
the charge ...”
According to what was referred by one of the its inventors, George E. Smith, it was originally in-
spired by magnetic bubble memories. The bubble technology was a shift register type structure in
which the presence or absence of a magnetic dipole represented 0s and 1s and the stored information
could be moved around using an appropriate magnetic eld. At the end of the 60’ the management at
Bell Labs demanded that a similar device should be devised using semiconductor technology. The
method which came to mind was the MOS capacitor in depletion regime. Charge could be introduced
into this depletion region, with the amount of charge stored without signi cant losses being the
magnitude of the signal. The last problem was how to shift the charge from one site to another,
thereby allowing manipulation of the information. This was solved by simply placing the MOS capaci-
tors very close together in order to pass the charge easily from one to the next by applying a more
attractive voltage to the gate of the receiver. The rst integrated device was an eight-bit shift regis-
ter. Since then, it did not take a long time to make an area imaging device, which is nowadays the
only practical application of CCDs. Nonetheless, it took many years to reach a massive spread of digi-

163

ff

fi
fi

fi

fi

tal imaging into the consumer market. In fact, this goal was achieved only after a dramatic improve-
ment in performance and a huge reduction of cost.
In this section, we will rst describe how the inversion charge can be stored and manipulated in
MOS capacitors. We will then address important practical issues, such as: the construction and opera-
tion of the output stage of a CCD; the limits on charge storage; and the mechanism of charge trans-
fer. Finally, the application of CCDs to image recording will be discussed in detail.

The charge storage paradigm


Figure 4.15 A) shows a single MOS capacitor biased with a positive gate voltage above threshold.
As we already discussed in Section 4.2, if the gate voltage is rapidly swept from accumulation to
VG > VT, the MOS capacitor will be biased into deep depletion, since there is no su cient time for
the inversion charge to build up. The corresponding band diagram is also shown in Figure 4.15 A).
Since this is a non-equilibrium state, the electron and hole quasi-Fermi levels Fn and Fp, are not equal
to each other. Note that Fn is close to the valence hand, indicating that the density of electrons is
quite small.
Figure 4.15 B) shows the MOS capacitor after a long time has elapsed and equilibrium has been
re-established. The depletion region has decreased in thickness to WT, namely has reached the size at
threshold for inversion. The surface potential has reduced to ϕsinv and the voltage drop across the
oxide has increased due to the inversion charge (observe the change in slope of the oxide bands).
When the equilibrium has been reached, the quasi-Fermi levels Fn and Fp merge into a single Fermi
level EF.
The application of MOS capacitors for image recording requires a very low rate of charge genera-
tion from thermal energy, in such a way that a long time of seconds or even minutes at room temper-
ature is required to ll the potential well. This time is long enough to allow MOS capacitors to store

Figure 4.15 MOS capacitor prepared in deep depletion


(A) and after equilibrium has been restored through ac-
cumulation of depletion charge (B).

164

fi
fi

ffi
the electron charge generated by the absorption of photons. Further, the stored charge can be manip-
ulated by transferring it from each gate to the adjacent one as we will see in the next sub-section.

Transfer of Charge between MOS Gates


Consider two MOS capacitors in close proximity, su ciently close for their depletion regions to
overlap, as shown in Figure 4.16. First, suppose that the capacitor on the left side is in inversion
(VG1  >  VT) with a stored charge, while the one on the right side is biased in depletion, but with
VG2 < VT. It is easy to see that the electrons are con ned beneath the capacitor on the left because
ϕs(lef t) > ϕs(right) and thus there is a lateral electric eld at the interface: ℰx = − d ϕ /d x > 0 (Figure
4.16 A). The surface potential is also plot in Figure 4.16, with an inverted axis. Then, if the gate po-
tential VG2 is raised to VG1, the electric eld vanishes and the charge will be shared equally between
the two gates (Figure 4.16,B). Finally, if we decrease the voltage on the left gate, the charge moves to
the right gate and there remains (Figure 4.16,C). Hence, we have transferred the charge from one
gate to another. It is important to realise that the initial inversion charge need not be the equilibrium
inversion charge. It is, generally speaking, a charge smaller than this, and possibly even zero charge.

Figure 4.16 Charge con nement in presence of two gates that share the
same bulk.

In (Figure 4.16, we have plotted the surface potential, with an inverted axis because this leads to
a diagram that is particularly useful for visualising the transfer and con nement of the inversion
charge. Because there is an electric eld pointing from regions of large ϕs to regions of small ϕs, the
large ϕs regions form a potential well for electrons that have been pictorially represented as small
dots in Figure 4.15 and Figure 4.16. To be consistent with this picture, it is convenient to visualise
electrons as moving like a uid that ows to adjacent wells when the potential barriers are removed.
Actually, this analogy can be carried much further, as illustrated in Figure 4.17 B.
The CCD shown in this diagram is known as a “three-phase CCD” and is the one originally pro-
posed by the inventors to demonstrate the principle of charge transfer in this family of devices. The
CCD is driven by three clock signals (or phases) ϕ1, ϕ2 and ϕ3 (Figure 4.17 A), which correspond to
the gate voltages VG1, VG2, VG3 and should not be confused with the interface potential ϕs, also depict-
ed in Figure 4.17 B. Note that all three clock signals are never high at the same time, and thus there
is always a gate with small ϕs to form a barrier between the potential wells in which charge packets
are stored.
Referring to the diagram for t  =  t1, only the clock signal ϕ1 is high and thus, charge packets are
con ned beneath the ϕ1 gates. In the diagram, from left to right, we have a nearly full, a partially full,
and an empty well under di erent gates, driven by the same phase potential ϕ1. These charges repre-
sent the signal to be transferred. At t = t2 the clock phase ϕ2 goes high and the charge packets are
shared between ϕ1 and ϕ2 gates. Note that, just as would be the case for a uid, the potential wells

165
fi

fi
fl
ff
fi
fl
fi
fi
ffi
fi
fi
fl

are less lled when charge is shared between two gates. At t = t3, ϕ1 begins to decrease and thus
charge ows from ϕ1 gates to ϕ2 gates. Finally, at t = t4, the charge is con ned beneath the ϕ2, gates
only. As time goes on, with the next clock steps, the charge packets are transferred to ϕ3 electrodes
and nally back to the next ϕ1 electrodes at the completion of a clock period T.
For a CCD of this type, it is clear that we can con ne one charge packet for every three gate elec-
trodes, so the information carrying capacitors are one-third of the total:

nin fo = ngrates /3. (4.5.1)

Since a charge packet moves three gates in a clock period T, the output signal is delayed by a time:

tdelay = T ngates /3 = ngates / (3 fclock), (4.5.2)

where fclock = 1/T is the frequency of the clock. With suitable input and output stages, such a CCD
could in fact, be used as a delay line, according to the original design mimicking magnetic bubbles.
Actually, nowadays the only practical use of CCD is image acquisition and a 3-phase clock is sel-
dom used in modern devices, which rely on either a 2-phase or a 4-phase clock. However, these
charge transfer schemes require a di erent architecture that will be addresses in the next sub-sec-
tions.

Figure 4.17 Mechanism of charge transfer in a three


phase CCD,

Before discussing image applications, we will rst consider the con nement of charge beneath a
gate electrode. We are concerned not only with the limit imposed by charge over ow beneath adja-

166
fi

fl
fi


ff
fi
fi
fi

fi

fl

cent electrodes sharing the same bulk, but also about con nement in the direction transverse to the
direction of motion of charge packets.

Charge Con nement beneath the MOS Gate


Consider an MOS capacitor with a particular inversion charge density Qn [C/cm2] that is either
equal to or less than the equilibrium inversion charge. Figure 4.14 shows that the interface potential
ϕs decreases with the inversion charge when the gate voltage VG is held xed. Therefore, in order to
nd the maximum charge that can be stores beneath a gate we can start considering Eq. (4.4.33) that
was already used to calculate the maximum charge that can be accumulated in an isolated MOS ca-
pacitor. We remind that the maximum inversion charge Qn is the one that brings the potential ϕs to
the threshold level ϕs , because at this point the potential well starts becoming leaky, like a forward
inv
biased diode. We call this charge retention capability “vertical con nement” to indicate that this limit
does not consider any interaction of the capacitor with other capacitors in the vicinities. To provide a
quantitative estimation of maximum Qn, it is convenient to recall the most relevant equations already
discussed in section 4.2:
The equation for vertical con nement is:

( )
Qn ma x = − Ci VG − ϕsinv − 2 V0 ϕsinv , (4.5.3)

where:

( ni )
kT NA
ϕsinv = 2 log and V0 = qεs NA /Ci2 (4.5.4)
q

By substituting Eqs. (4.5.4) into eq. (4.5.3), we easily get:

( ni ) ti [ q ( ni ) 2 ]
Na ϵ kT Na 1
Qn ma x = − 2 ϵs Na k T log + i log − VG (4.5.5)

To estimate the maximum charge QnAma x , where the superscript “A” refers to the speci c area of the
device under evaluation, we consider a MOS capacitor characterised by the same parameters used to
calculate the quantities listed in Table 4.1, with the addition of realistic values for the gate area A and
the bias voltage VG:

Table 4.2. Parameters of a silicon based MOS capacitor

Parameter εi εs Na ti VG A T

Value 3.45 ×10-11 F/m 1.05 ×10-10 F/m 1022 m-3 50 nm 10 V 10-10 m2 300 K

For a MOS capacitor with an area A = 10 x 10 μ m 2 we get QnAma x ≈ 6 × 10−13 C , which corresponds
to a number of electrons: Ne ≈ 3 × 106. This gure is de nitely larger than the charge can be stored
in real devices of comparable active area. In fact, practical operation of a CCD imposes additional
constraints that limit the charge to a lower value. Let us consider rst the requirement that the

167
fi

fi

fi


fi

fi

fi
fi
fi

fi
fi

charge remain con ned to a gate with VG1 = VGH when the adjacent gate is biased in inversion regime
with a voltage below threshold (VG2 = VGL). This situation is depicted inFigure 4.18.
The charge is con ned beneath the left-hand gate if ϕs1 > ϕs2, which is the condition for an electric
eld to establish in the region between the two gates. Conversely, the charge will begin to be shared
between the two gates when over lling brings the surface potentials to become equal. Therefore, the
maximum charge that can be stored beneath the left gate can be calculated from Eq. (4.4.30), written
twice to represent the interplay of the two MOS capacitors, which have reached the same potential ϕ̄s
due to the charge accumulated in one of them, upon di erent bias voltage VGH and VGL:

Qn ma x
VGH = ϕ̄s + 2V0 ϕ̄s − Ci
(4.5.6)
VGL = ϕ̄s + 2V0 ϕ̄s

The maximum inversion charge allowed by this “lateral con nement” mechanism is then easily found
by taking the di erence of Eqs. (4.5.6):

Qn ma x = − Ci (VGH − VGL) (4.5.7)

From equation (4.5.7) we understand the importance of a proper design of the capacitor in terms on
thickness of the insulator and choice of the dielectric material. In fact, the maximum charge that can
be stoppered bene ts from a thin insulator and a large dielectric constant. Using the parameters list-
ed in Table 4.2, the maximum charge that can be stored for the given gate bias voltages results to be
QnAma x = 3 × 10−13 C, or equivalently: Ne ≈ 2 × 106.
Generally speaking, a large Qn ma x is desirable, as this improves the dynamic range of the detector
and the signal to noise ratio. This suggests maximising VGH and minimising VGL. However, as VGH is
increased, the electric eld in the depletion region increases, and eventually avalanche breakdown
will occur. VGL cannot be set to zero because in the absence of any depletion layer charge trapping in
interface states may occur, which leads to degraded transfer e ciency. Usually VGL is chosen to pro-
duce just a depleted surface, and VGH is typically determined by system constraints.
An Interesting consequence of Eq. (4.5.7) is that the maximum stored charge is independent of
doping concentration. This is true despite the fact that the potential well is "deeper" (namely,
ϕs[Qn = 0] is larger) for smaller Na. Actually, for lightly doped substrates the potential well is initially
deeper, but it lls up more rapidly.
The above discussion refers to the con nement along the direction of charge transfer and can be
switched ON and OFF by the interplay of the gate voltages. Next, we consider con nement trans-
verse to this direction. Typically, con nement is obtained by increasing the oxide thickness near the
gate edge, or the doping concentration (known as a channel stop), or both. We consider here the ef-
fect of an increased oxide thickness, as shown in Figure 4.19.
Suppose that the gate voltage is high (VG = VGH), while a charge Qn ma x is stored beneath the thin
oxide of thickness ti1. Then, the surface potential is the solution of the equation:

Qn ma x = − Ci1 (VGH − ϕs1 − 2 V01 ϕs1 ), (4.5.8)

2
where Ci1 = ϵi /ti1 and V01 = qεs NA /Ci1 . The maximum charge corresponds to the situation when
the charge is just barely con ned beneath the thin oxide. That is when we have a surface potential ϕs2
beneath the thick oxide equal to ϕs1, but no inversion charge. Namely:

168
fi
fi
ff
fi
fi
fi
fi

fi

fi

fi

fi

ff
fi

ffi

fi

Figure 4.20 Schematic of the output stage of a Figure 4.21 Output stage of a CCD showing the
CCD with a MOSFET biased as voltage follower. MOSFET that generates the potential well beneath
the n+ diffusion.

Figure 4.22 Pictorial representation of the readout


process in a CCD.

169
Figure 4.18 Charge con nement along the direc- Figure 4.19 Charge con nement produced by an
tion of charge transfer. increase in oxide thickness.

0 = − Ci2 (VGH − ϕs1 − 2 V02 ϕs2 ) with ϕs2 = ϕs1. (4.5.9)

The simultaneous solution of (4.5.8) and (4.5.9) yields the maximum charge that can be stored un-
der the left-hand side of the gate. This can be made larger than the limit in (4.5.7) if ti2 is chosen suf-
ciently greater than ti1.

The Output Stage of a CCD


After discussing charge con nement and transfer we need to address the extraction of information
from a CCD through an output stage that must convert stored charge into voltage. A circuit that per-
forms this function is illustrated in Figure 4.20. The charge is transferred to a capacitor Cout that is
also connected to the gate of a MOSFET biased as a voltage follower.
A voltage follower circuit can be implemented either with a Bipolar Junction Transistor (BJT) or
with a Filed E ect Transistor (FET). In case of a FET, the Source Follower replicates to the source
(low impedance) the input voltage set to the gate (high impedance), except for a quasi-constant
voltage drop (VGS) required to keep open the channel. Here the follower is implemented with a MOS-
FET, since the voltage to be replicated is obtained by charging the MOS capacitance of the transistor.
Assuming that the capacitor Cout is linear, namely its capacitance is not voltage dependent, the voltage
VC across the capacitor is:

QC
VC = , (4.5.10)
Cout + CG

where QC is the charge transferred to the capacitor Cout, and CG is the input capacitance of the MOS-
FET. When the MOSFET is biased with a current generator, as shown in Figure 4.21, its output volt-
age Vout is:

Vout = VC − VGS, (4.5.11)

where VGS is the constant gate-to-source voltage required to bias the transistor to the drain current
Ibias.
The implementation of the output stage is shown in Figure 4.21. The output capacitor is the deple-
tion capacitance of the n+ di usion. Also illustrated in the gure is a second MOSFET that
“precharges” the di usion to a positive voltage +VDD. The negative inversion charge QnA = QC, which
represents the signal, is transferred from the adjacent gate, reduces the amount of reverse bias on the

170
fi

ff


ff
fi
fi

fi
ff

fi

output capacitor Cout, and thus change the input signal to the voltage follower. Note that the inver-
sion charge, which has been indicated with symbol QnA, is now a “real” charge obtained from the in-
tegral of the inversion charge density Qn over the active area A of the MOS capacitor.
More in detail, the operation of the CCD output stage is illustrated in Figure 4.22. To begin with
Figure 4.22 A), the di usion is precharged to the voltage +VDD through the control potential ϕR;
both phases ϕ2 and ϕ3 are high, and the inversion charge is shared between the ϕ2 and ϕ3 gates. In
Figure 4.22 B) ϕ2 goes low and the charge is shifted beneath the ϕ3 gate alone. In Figure 4.22 C) ϕ1
is high, and the charge is shared between the ϕ2, ϕ3 gates and the di usion n+. As rst ϕ3,(Figure
4.22 D) and then ϕ1 go low, the charge is pushed to the right, so that it is all in the di usion poten-
tial well (Figure 4.22 E), thus changing the voltage sensed by the MOSFET voltage follower.
An approximate analysis of the output stage leads to the following expression for the voltage
across the output capacitor:

QnA
VC ≈ VDD + , (4.5.12)
CG + CD

where CG is the input capacitance of the MOSFET and CD is the average value of the depletion capaci-
tance over the output voltage range.
Actually, modern CCDs use a more re ned output stage called Floating Di usion Ampli er (FDA)
As shown in Figure 4.23, the FDA consists of a node for detecting charges and two MOSFETs
(MOS1 for reset and MOS2 for charge-to-voltage conversion) connected to the node. The charge
transferred to the detection node is converted into a voltage by MOS2 via Eq. (4.5.10). The detection
node is reset by MOS1 to the reference level (voltage on RD) in order to read the next signal. Noise
accompanying the charge detection by FDA is caused by the capacitance of the node, but can be al-
most entirely eliminated by using a methods named Correlated Double Sampling (CDS). CDS elimi-
nate or at least reduce the variation of the common mode of the ampli er and the reset noise of the
oating node.

Figure 4.23 Schematic of the Floating Diffusion Am-


pli er in the output stage of a CCD.

171
fl
fi


ff

fi
ff
fi
ff

fi
ff
fi

Realistic schemes commonly used in CCDs


Modern CCDs are highly complex devices, whose analysis is far beyond the scope of the present
notes. Nevertheless, some common implementations must be mentioned before considering the gen-
eral architecture of image sensors.

FOUR PHASE CLOCK CCD

Figure 4.24 shows the schematic of a four-phase-clock CCD. Starting from a p-type substrate, a
rst set of poly-silicon (conductive and semi-transparent) gates are made over an oxide layer. Then,
the exposed surfaces of the poly-silicon gates are oxidised. A second deposition of doped poly-silicon
is performed, and a second set of gates are de ned that overlap the rst poly-silicon. This results in a
gap between gates that is equal to the thickness of the oxide layer. In this structure, it is desirable for
all oxide layers to have the same thickness in order to maximise the charge that can be stored. The
four-phase CCD is less dense than other designs, but has high-charge storage capacity and a simple
design. The charge transfer mechanism in a four phase clock CCD is shown in Figure 4.25

Figure 4.24 Schematic of a four phase clock CCD. Figure 4.25 Charge transfer mechanism in a four
phase clock CCD (from Hamamatsu Photonics).

TWO PHASE CLOCK CCD

In the two-phase CCD, unequal oxide thicknesses are used to make a simpler structure. This CCD
implementation requires only two clock phases and two gates per stage. The cross-section of a two-
phase CCD is shown in Figure 4.26. A doped poly-silicon layer is etched to form the rst set of gates
(grey in gure). Then, an oxidation step is performed, which intentionally creates an oxide layer of
unequal thickness. Contact holes are etched and a second set of gates is deposited and de ned. Note
that the rst and second gate conductors are connected to form a single gate. This CCD structure is
driven by two-phase, non-overlapping clock signals (that is, ϕ1 and ϕ2 are never high at the same
time). Transfer of charge is illustrated in Figure 4.26. More speci cally, Figure 4.26 B) shows the
surface potential when ϕ2 is high and ϕ1 is low. From Eq. (4.4.31) we know that the potential well ϕs
with zero charge is given by:

172
fi
fi
fi

fi
fi
fi
fi
fi

2 VG
ϕs = ϕs (Qn = 0) = VG − V0 1+ −1 , (4.5.13)
V0

where V0 = qεs Na /Ci2 . Thus, the potential well is deeper below the thinner oxide, since the gate
with larger Ci has lower V0 and more positive ϕs. So, the charge packets will reside beneath the thin
oxide under the ϕ2 gates. In Figure 4.26 C) we see that when ϕ2 is low and ϕ1 is high, the charge is
transferred to the right. In order for this design to work properly, the charge must never be large
enough to "over ow" beneath the portion of the gate with the thicker oxide. This means that the
maximum stored charge cannot be as large as for a three or four phase CCD. While the charge capaci-
ty is limited, this design is particularly simple and dense.

Figure 4.26 Schematic of a two-phase CCD.

BURIED-CHANNEL CCD

It is known that the region of interface between a semiconductor and another material is a ected
by a high density of defects. Interface states in semiconductors act as generation/recombination cen-
tres and can trap carriers. Therefore, the simple scheme of MOS capacitor that we have seen so far is
not the best option when the rate of thermal generation of couples e-h has to be kept as low as possi-
ble and the transfer of electrons must occur without any loss. The buried-channel (or bulk-channel)
CCD is a variant motivated by preventing trapping and emission of electrons from interface states. By
con ning the charge to a potential well that is located away from the interface, the in uence of inter-
face states is totally eliminated. To see how such a potential well can be formed we start from Figure
4.7 A) that also shows the band diagram of an MOS capacitor biased into deep depletion. Here the
maximum electric eld in the semiconductor is at the insulator-semiconductor interface, where the
potential well is con ned. By using the Gauss' law, we have already seen that:

qNaW
ℰs = , (4.5.14)
ϵs

where W is the width of the depletion region. Suppose now that we insert an n region at the surface
and simultaneously readjust the gate voltage to keep the thickness of the depleted p region un-

173
fi

fl
fi
fi

fl

ff
changed (Figure 4.27 A). Assuming that the n region is completely depleted, this is equivalent to
inserting positive charges (donor ions) near the interface, and thus a smaller gate voltage is required.
In fact, the separation between the Fermi level of the metal Fm and the quasi-Fermi level for holes Fp
is larger in Figure 4.7 B) than in Figure 4.27 A), which also shows the change in the shape of the
bands that produces the buried channel. In the p region, the ionised acceptors create a negative space
charge, which causes the bands to bend downward; the n region contributes a positive space charge,
so the bands must bend upward toward the surface of the insulator. If the thickness of the n region is
large enough, the electric eld reverses sign at the surface and a potential well for electrons is formed
in the semiconductor bulk. As electrons are added to the potential well, the additional negative
charge causes the thickness of the depleted p region to decrease as shown in Figure 4.27 B). Eventu-
ally, the electron concentration becomes equal to the donor concentration, and a portion of the n re-
gion becomes electrically neutral, which corresponds to a null electric eld. When this takes place,
the depletion regions in the n and p materials cease to interact, and we have an n-MOS capacitor in
series with a pn-junction diode (not shown). This sets the limit to the electron charge that can be
stored in the buried channel. As noted above, the major advantage of the buried-channel CCD is the
absence of interface trapping, which permits extremely high charge transfer e ciency in term of elec-
tron loss. This is particularly valuable for CCD imagers. The maximum stored charge is reduced,
however, compared to surface-channel devices with the same oxide thickness, and the process com-
plexity is increased.

Figure 4.27 Buried channel MOS capacitor.

The CCD image sensor


The development of the rst image detector based on a CCD dates back to a short time after the
invention of this family of devices. This application was pushed by the interest of building a video-
telephone service that, at that time, was called “Picturephone5” (Figure 4.28). Since then, a long
time has passed and an impressive technological advancement has brought to the development of
highly performant devices. Figure 4.29 shows the complexity of the element of a modern CCD (from
Hamamatsu Photonics).
CCDs for imaging applications can be divided into two main types: linear arrays and area arrays.
Linear arrays consist of one, or more generally few lines of light sensitive pixels, while area arrays
consist of matrices of light sensitive pixels. Linear arrays are used in scanners and spectroscopic in-
struments. Area arrays, beyond in consumer electronics, are much more common for scienti c and
high end imaging. Therefore, the remainder of this section will focus on these devices.
Di erent architectures for area array CCD are depicted in Figure 4.30. Figure 4.30 A) shows the
schematic of a full frame CCD. The entire area, corresponding to all the MOS capacitors, is used to

5For reference consider the Stanley Kubrick’s movie 2001: A Space Odyssey released on April 1968!

174
ff

fi
fi
fi
ffi

fi
Figure 4.28 Picturephone (1968). Figure 4.29 Pictorial schematic of an element of a
modern CCD detector.

collect light. This provides the optimal exploitation of silicon. Upon completion of the exposure, the
rows of the sensitive matrix are transferred one-by-one to the readout register at the bottom row of
the matrix. After each row has reached the readout register, the charge accumulated in each pixel is
shifted one packet at a time to the end of the row, where it is converted to a voltage, as discussed in
section 4.4, and eventually digitised. Since the readout process is rather slow, full frame CCDs are not
suitable for high speed acquisition and require a camera shutter to close during readout, so that elec-
trons are not generated when charge is transferred, which would result in image streaking. Figure
4.30 B) shows the schematic of a frame transfer CCD. The sensor is made by two equal matrices of
pixels, one is exposed to the light and collects photons during the integration time, the other (frame
store area) is covered with an opaque mask. At the end of the exposure, the charge accumulated in
the sensitive matrix is rapidly transferred to the frame store area. A mechanical shutter might or
might not be required according to the time required for the charge transfer, compared to the light
integration time. In frame transfer CCDs, integration can begin for the next image frame as soon as
the frame transfer is complete and while the frame store section is reading out. In these devices the
image acquisition and the frame store clocks must be driven separately, requiring slightly more com-
plicated electronics than for full frame devices. Yet, this con guration allows a higher frame rate than
full frame CCDs. Further, the pixel readout rate can be made small to reduce one of the major
sources on noise in CCDs, as we will see in a short time. The main disadvantage of frame store de-
vices is that only half of the silicon area is light sensitive. Full frame and frame transfer CCDs are typ-
ically used in scienti c and high end applications when cost is not a major concern. Figure 4.30 C)
show the principle of Interline transfer CCDs. They have a shift register masked from light alongside
each column of photosensitive pixels so that the charge accumulated in the elements of the imaging
matrix can be transferred very rapidly to the shift registers. The pixels in the shift registers can then
be read out while the next image frame is being exposed. This allows for a faster frame rate and no
external shutter is required. Interline transfer devices are more complex (in terms of fabrication) than
the other types but are commonly used in high speed imaging system and in consumer devices,
where a mechanical shutter would not be practical. They are seldom used for high end scienti c
imaging systems since the ll factor is rather poor. Sometimes an array of microlenses is placed over
the sensor in order to focus the light onto the sensitive region of each pixels. Finally, Figure 4.30 D)

175
fi
fi
fi
fi
Figure 4.30 Different architectures for CCD based image detec-
tors.

shows the schematic of an Orthogonal Transfer CCD (OTCCD). In these sensors the channel stop
that prevents charge loss along the direction orthogonal to the shifting direction is replaced with an
actively clocked phase, so that charge shifting in both directions (along rows and columns) may be
achieved. If centroiding of a moving object in the scene is performed with another detector, the feed-
back can be used to clock the OTCCD in any direction to minimise image blurring. This is a useful
function when making astronomical observations in which atmosphere motion (scintillation) blurs
images, unless a mechanical tracking system (e.g. a deformable mirror) redirects the optical beam
based on a feedback signal.

SPECTRAL SENSITIVITY

The spectral sensitivity of silicon based devices depends on two major factors: i) in the short wave-
length range photons are absorbed in a very thin layer, which might not correspond to the depletion
region where charge accumulation takes place; ii) conversely in the far red and NIR spectral range the
absorption length might be larger that the thickness of the depletion layer, thus limiting sensitivity.
Figure 4.31 depicts the expected responsivity of a CCD as a result of the above mentioned e ects.
To signi cantly improve the sensitivity, mainly in the short wavelength range, the bulk of high end
image sensors can be thinned to few tens of microns, in such way that the light can enter the device
from the back side. These detectors, called “Back-illuminated” CCDs (Figure 4.32)require additional
post fabricated steps, which adds considerable cost. However, back illuminated devices are much
more e cient in detecting light than front-illuminated and usually have high sensitivity throughout a
very broad spectral range. For many highly demanding applications, back-illuminated (or back-
thinned) CCDs are the detector of choice even though they are more expensive than their front-illu-
minated counterparts. Quantum E ciency (QE) is limited only by re ection at the back surface and
the ability of silicon to absorb photons, which is a function of device thickness and operating temper-
ature. An antire ection coated back illuminated CCD may have a peak QE > 98% in the visible. Op-
tical absorption and multiple re ections from frontside structures are avoided with backside devices
although there is still interference fringing due to multiple re ections within the thin CCD. This in-
terference fringing is often worse for backside sensors than it is for frontside sensors. In recent years

176
ffi
fi
fl

fl
ffi
fl
fl
ff

Figure 4.31 Expected responsivity of silicon based


CCDs.

many back illuminated sensors have been made much thicker than previously possible to increase
absorption in the red and to decrease the amplitude of fringing.
Figure 4.32 shows the typical quantum e ciency of front-illuminated and back-illuminated CCDs.
Front-illuminated CCDs have no sensitivity in the ultraviolet region, and their maximum quantum
e ciency in the visible region is approximately 40%. In contrast, standard back-thinned CCDs pro-
vide very high quantum e ciency, which can be 40% or higher in the ultraviolet region and approxi-
mately 90% at a peak wavelength in the visible range. The fully-depleted back-illuminated CCDs use
a thick silicon substrate which allows higher sensitivity in the wavelength range from 800 to 1100
nm than standard back-thinned CCDs.

Figure 4.32 Schematic of a Front-Illuminated versus a Back-Thinned


Frame-Transfer CCD.

The sensitivity in the visible region from 400 to 700 nm can be improved with an anti-re ection
coating; however the ultraviolet sensitivity is, in general, rather low. The spectral response range is
determined on long wavelengths by the silicon substrate thickness and on short wavelengths by the
sensor structure on the light input side. Front-illuminated Frame-Transfer CCDs require that poly-
silicon gate electrodes be formed on the e ective photosensitive area because of their structure,

177
ffi
ffi

ffi
ff

fl
which makes the CCD almost insensitive to ultraviolet light at wavelengths shorter than 400 nm. To
give sensitivity to ultraviolet light, a scintillator material called “Lumogen” is coated on the CCD sur-
face of some front-illuminated detectors. Back-illuminated CCDs, on the other hand, provide a high
quantum e ciency from ultraviolet to near infrared and also exhibit excellent stability for prolonged
exposure to ultraviolet light (Figure 4.33).

Figure 4.33 Typical quantum ef ciency of front- and


back-illuminated CCDs.

DARK CURRENT

As already seen in all the quantum detectors discussed so far, also in image sensors spontaneous
generation of couples e-h gives rise to a dark current in absence of any input light. The dark current is
generally expressed in units of e-/pixel/s, which indicate the number of electrons generated in one
pixel per unit time. Dark current nearly doubles for every 5 to 7 °C increase in temperature and con-
tributes to the noise in CCDs, as we will better see shortly. The three major causes of dark current in
CCDs are:
Thermal excitation in depleted layer
Thermal excitation in undepleted region and di usion to the depleted layer
Thermal excitation by surface levels
The dark current from a silicon pixel can be model by the semi-empirical expression:

[s]
Eg
( kT)
e
id = 2.5 × 1015 Api x DFM T 1.5 exp − , (4.5.15)

where DFM is a silicon ‘quality’ parameter, expressed in nA/cm2 @300K, Apix is the pixel area, Eg is
the silicon band gap, T is temperature, and k is the Boltzman constant.
The usual method adopted to reduce dark current requires cooling the detector. If very low light
level measurements have to be made, like in astronomic observation (spectroscopy or imaging), it is
common to cool a CCD to −100 °C or to an even lower temperature. This reduces the dark current to
just a few electrons per pixel per hour, but also reduces the quantum e ciency. For higher light level
measurements, devices cooled to −40°C with a thermoelectric cooler are commonly employed. Many

178

ffi
fi

ff

ffi

commercial CCDs (or CMOS sensors) operate with no cooling. In this case the dark current might be
so high that the exposure time must be rather short (few seconds or less) and quantitative measure-
ments are critical since the dark electrons overwhelm the charge generated by the incident light.

NOISE IN CCDS

The noise that a ects CCDs and other images sensors, like CMOS sensors, can be classi ed into
the following four main sources:

Fixed Pattern Noise (Nf)


This noise is caused by variations in sensitivity between pixels. These variations may be due to QE
di erences, non-uniformities in the aperture area and in the thickness and instabilities in electronics
which are devoted to image readout. Images can be ‘ at eld corrected’ by dividing data images by a
calibration image taken with uniform illumination. Yet this correction increases random noise in the
resultant image.

Shot Noise of the Signal (Ns)


Shot noise is the noise generated by statistical variation in the photon ux incident on a CCD.
Shot noise is expressed by the standard deviation of the number of photons collected per pixel during
the exposure time. Assuming that the incoming ux of photons obeys the Poisson distribution, the
shot noise is given by the well known expression for the standard deviation of a Poisson process:

Ns = S, (4.5.16)

where S represents the Signal, namely the number of light generated electrons in a pixel in the expo-
sure time. Being this noise source fundamentally related to the quantum nature of light, its contribu-
tion is unavoidable.

Dark Shot Noise (nd)


Dark shot noise is caused by dark current and is proportional to the square root of the number of
electrons generated per pixel in dark state. The dark current largely depends on the quality of silicon
and can be e ectively diminished with cooling. To further reduce the dark shot noise the acquisition
time must be limited.

Readout Noise (Nr)


The readout noise is mainly due to electrical uctuations from thermal noise caused by the MOS-
FET used as the ampli er in the CCD output stage. It also comes from the readout circuit of the
CCD. The Readout noise is the fundamental uncertainty in the output of a CCD; it is the dominant
noise source in high performance imaging systems and eventually sets the lowest detection limit. The
readout noise is typically measured in electrons rms, but is actually a voltage uncertainty. It is not af-
fected by the amount of exposure, but is determined by the CCD output and by its operating speed,
which sets the bandwidth of the system. As the readout speed increases, the readout noise increases
as well by a square root of the frequency, as expected for thermal noise, whose PSD is almost at up
to a very high frequency. At low readout frequency, the readout noise is dominated by 1/f or icker

noise. Figure 4.34 shows the dependance of the readout noise, expressed in erms , versus frequency,
for a typical CCD. A more detailed discussion of the readout noise will be presented in section 4.6.

Overall Noise Figure


As it is appropriate for independent noise sources, the way to combine the di erent noise contri-
butions into the total noise Nt requires to take the square root of the sum of the square noise terms:

179
ff

ff

ff

fi

fl
fl
fl
fi
fl
ff

fi
fl
fl

Figure 4.34 Readout noise as a function of the frequency of


the readout circuit in a typical CCD.

Nt = Nf2 + N s2 + Nd2 + N r2 . (4.5.17)

Figure 4.35 schematically shows the interrelation between the mentioned four noise terms versus
the amount of exposure in a logarithmic plot. The CCD detection limit is typically determined by the
dark shot noise and the readout noise, being the second more relevant in scienti c grade sensors.
Actually, the rst term can be controlled by cooling the detector, while the second term improves
when the frame rate, namely the operating speed, is reduced. The S/N is mainly determined by the

Figure 4.35 Signal and noise terms as a function of the exposure in a typical
CCD.

180
fi

fi

xed pattern noise when the amount of exposure is high, and it is determined by the readout noise or
the shot noise, depending on the device, when the amount of exposure is low.

DYNAMIC RANGE

The dynamic range speci es the range of light intensity measurable by a detector and is de ned as
the ratio of the maximum level to the minimum level in the output. The dynamic range should not be
confused with the depth in bits of the Analog to Digital Converter that produces the recorded image.
The CCD dynamic range can be obtained by dividing the saturation charge by the readout noise, as-
suming that this is the main noise source, as it it the case in scienti c grade sensors.

Saturation charge
Dynamic range = (4.5.18)
Dark noise
Alternatively, the Dynamic range can be expressed in logarithmic unit:

Saturation charge
Dynamic range = 20 × log [dB]. (4.5.19)
Dark noise
The dynamic range varies with operating conditions such as temperature and integration time. At
room temperature, the dark shot noise typically determines the lower detection limit and the dynam-
ic range. When su cient cooling is applied, in such a way that the dark shot noise can be ignored,
the readout noise determines the dynamic range.

AMPLIFIED CCD DETECTORS

In scienti c applications of CCDs, the ultimate limit to sensitivity in very low light conditions is
set by the shot noise of the photon ux. Nevertheless, the achievement of this quantum limit re-
quires the substantial suppression of the dark current and to accumulate the signal for a time long
enough to overcome the readout noise.
This approach is unpractical when studying dynamic phenomena, as often occurs in microscopy
when dealing with biological specimens. In that case the readout noise invariably set the sensitivity
limit. In order to overcome this constraint, an ampli cation method that operates before the conver-
sation of the charge packets into voltage has been devised. This stems from the general rule that re-
quires the ampli cation to operate before the noise enters into the system.
The device that exploits charge ampli cation to make the readout noise negligible is called “Elec-
tron Multiplier CCD (EM-CCD)”. Figure 4.36 shows the structure of an EM-CCD. The basic feature
is the same as a normal Frame Transfer CCD, typically in a back-thinned version. The charge accumu-
lated in the image area is rapidly transferred to the storage area, and then the charge stored in the
frame is transferred line by line to the horizontal serial register for readout, just as in a normal
Frame-Transfer CCD. At this point, before reaching the reading stage, in an EM-CCD the charge
packets are ampli ed in a Charge Multiplication Register built into the horizontal serial register.
Within this register, signal multiplication is done by supplying a voltage higher than normal to each
horizontal transfer electrode. Figure 4.36 shows the principle of signal multiplication in the charge
multiplication register. When the charge packet is transferred from stage to stage, the electrons are
accelerated by the high electric eld generated under the multiplication gate by applying a high volt-
age (30 to 40 V) to each multiplication electrode. This voltage is much greater than the normal hori-
zontal transfer voltage, and generates an occasional extra e-h pair. This occurs for the impact ioniza-
tion mechanism that has been already discussed to describe the signal ampli cation in avalanche
photodiodes. The probability of such an event is rather small, typically about 1.0 % to 1.6 % at each
stage. However, electrons are multiplied repeatedly in a large number of stages, so that a high gain is

181
fi
fi
fi
fi
ffi
fi
fi

fi
fl

fi

fi

fi

fi
Figure 4.36 Schematic that shows the operating princi-
ple of an EM-CCD.

achieved through the multiplication register. Normally, there are 400 to 600 stages in the multiplica-
tion register. The total multiplier gain M can be expressed by the following formula:

M = (1 + g )
N
(4.5.20)

where g is probability to generate an electron-hole pair at each stage and N is total number of charge
multiplication stages. The multiplication factor M can be changed from 1 to a maximum value (e.g.
Mmax = 100) according to the need.
The probability of generating secondary electrons depends not only on the supply voltage but also
on the temperature of the sensor. Control and stability of both the supply voltage and the CCD tem-
perature are very important factors when the EM-CCD technology is used for quantitative measure-
ments. The key advantage of an EM-CCD is that the signal charge in the CCD is multiplied before
being converted to voltage by the output ampli er. Multiplication of the signal before the charge to
voltage conversion makes the readout noise relatively smaller as the multiplication factor increases.
This happens because the readout noise of the output ampli er must be divided by the multiplication
factor M, in order to weight it for the ampli ed signal. Hence, even at moderate gain settings, the rel-
ative readout noise can become less than 1 electron, enabling detection of even single photon events
in each pixel. To observe low intensity signals at high speeds, the EM-CCD can overcome the growth
in readout noise with the frequency of the output ampli er, by additional multiplication gain, which
might still keep the relative readout noise to less than 1 electron. This ability to use the multiplica-
tion gain register to overcome the readout noise is the chief advantage of the EM-CCD cameras for
fast, low light imaging in scienti c applications. However, it must be kept in mind that the ultimate
shot noise, intrinsically present in the photon ux, is not a ected by the gain, which, on the other
hand, introduces extra noise due to the stochastic character of the impact ionisation process. There-

182

fi

fi
fl
fi
fi
fi
ff

fore, when acquisition speed is not a concern, gain should be applied with parsimony, otherwise im-
ages a ected by a severe granular noise, also called ‘salt and pepper noise‘ are recorded as a conse-
quence of the multiplication of the shot noise. Further, for quantitative imaging one must take into
account that the multiplication gain tends to su er from ageing. In fact, a slow decrease in gain over
time takes place, based on the total electric charge passed across the multiplication register. While
the exact cause of the gain degradation is not known, it is thought that the higher than normal volt-
ages used in the process leads to trapping of accelerated electrons in the transfer electrode. These
trapped electrons may change the electric eld and thus create the gain ageing phenomenon.
Figure 4.37 shows the improvement that can be obtained in the output image when the gain is
progressively increased, while keeping constant light. it is evident that unless a su cient electron
multiplication is applied, the signal is buried within the readout noise.

Figure 4.37 Improvement in the quality of an image taken


with an EM-CCD in low light condition for increasing val-
ues of the gain.

COLOR CCDS

The CCDs used in scienti c imaging are almost invariably native B&W detectors. Spectral resolu-
tion can be obtained using a set of passband lters or using more complex approaches based on dis-
persive elements, like gratings or acousto-optics modulators. Recently, hyperspectral cameras based
on interferometric methods to measure the full spectrum in each pixel of an image have been also
manufactured using either CCD or CMOS sensors. Conversely, in consumer electronics and in pro-
fessional photography color sensitivity is usually achieved at detector level by means of a Color Filter
Array (CFA) deposited onto the sensitive area, is correspondence of the pixel matrix. The most
common CFA used in color recording is the ‘Bayer lter mask’. This mask is a mosaic of RGB lters
that implements the additive model of color. In Bayer mask the fraction of green elements is 50% of
the total, which is double respect to red and blue elements. This design is compliant with the human
eye sensitivity, which is maximum in the green region of the visible spectrum, as seen in section 1.7.
Three separated R-G-B color planes are then recovered from the intensity map produced by the Bayer
mask using an appropriate interpolation procedure, called ‘demosaicing’. In many detectors, an ap-
propriate low-pass lter is combined with the Bayer mask to avoid aliasing caused by the under-sam-
pling of the optical eld in the three color planes. Figure 4.38 shows the schematic of the Bayer color
mask.

183

ff
fi
fi
fi
fi
fi
ff
fi

ffi

fi
Figure 4.38 Schematic of the Bayer lter color array.

184

fi
Section 6

CMOS VS. CCD IMAGE SENSORS


CCD and CMOS sensors are competing technologies for capturing digital images that share the
same operating principle based on the MOS capacitor. Yet, the management of the charge packets ac-
cumulated in pixels is fairly di erent. Each technology has unique strengths and weaknesses, giving
advantages in speci c applications.
In a CCD sensor the charge packets originating from photonics signals are transferred to a com-
mon output node, where the electric charge is converted to voltage, bu ered, and sent o -chip as an
analog signal. A large fraction of the sensor area can be devoted to light capture, and the output uni-
formity (a key factor in image quality) is high. On the other hand, most of the functions required for
digital imaging take place on the camera printed circuit board, which increases the cost of the whole
system. In a CMOS sensor, each pixel has its own charge-to-voltage conversion, and the sensor often
includes ampli ers, noise-correction, and digitisation circuits, so that the chip outputs digital bits.
These functions increase the design complexity and reduce the area available for light capture. With
each pixel doing its own conversion, uniformity is lower, but it is also massively parallel, allowing
high total bandwidth leading to high frame rate. Further, CMOS sensors can be operated with a single
power supply and provide more exible readout options compared to CCD, with easy implementa-
tion of Region-Of-Interest (ROI) or windowing that limit the data throughput. Figure 4.39 shows
the di erent architectures of a CCD and a CMOS Image Sensor (CIS).
CCDs are generally made in NMOS technology (p-type substrate) which is dedicated to imaging
performance and start from high quality material. CMOS sensors are often consumer oriented, based
on standard CMOS process technology for digital ICs, with some adaptation for imaging. It is gener-
ally considered that the manufacturing of CMOS sensors is cheaper than CCD and that the perfor-
mance is lower. This assertion is based on the extremely large volume of the mass market (e.g. cell
phones) that has pushed the development of CMOS sensors to the point that, nowadays, CCDs are
mainly restricted to scienti c applications. As an example, the large majority of space programs are

Figure 4.39 Schematic of a CCD and a CMOS Image Sensor (CIS).

185
ff
fi
fi
fi
ff
fl

ff
ff
still based on CCD components not only for the performance, but also to ensure the long-term sup-
ply – not always compatible with consumer demand. Similarly, the scienti c imaging market is also
still served by a majority of high-end CCD based solutions, with new product developments still in
progress.
Interest in CMOS started a long time ago, based on expectations of lowered power consumption,
camera-on-a-chip integration, and lower fabrication costs from the reuse of mainstream logic and
memory device fabrication. Achieving these bene ts in practice while simultaneously delivering high
image quality has taken far more time, money, and process adaptation than original projections sug-
gested, but CMOS imagers have surpassed CCDs as mainstream, mature technology, since several
years ago. As a matter of fact, CMOS designers focused e orts on imagers for mobile phones, the
highest volume image sensor application in the world. An enormous amount of investment was made
to develop and ne tune CMOS imagers and the fabrication processes. As a result of this investment,
we witnessed great improvements in image quality, even as pixel sizes shrank. Therefore, in the case
of high volume consumer area and line scan imagers CMOS are nowadays the only option. The per-
formance advantage of CMOS imagers over CCDs for machine vision, which is one of the largest in-
dustrial markets, is also established. For machine vision, the key parameters are speed and noise. In
CMOS area and line scan imagers, the front end of data path is massively parallel. This allows each
ampli er to have low bandwidth. By the time the signal reaches the data path bottleneck, which is
normally the interface between the imager and the o -chip circuitry, CMOS data are in the digital
domain. In contrast, high speed CCDs have fast output channels not as massively parallel as high
speed CMOS imagers. Hence, each CCD ampli er requires high bandwidth, which results in higher
noise. Consequently, high speed CMOS imagers can be designed to have much lower noise than high
speed CCDs. In application demanding high frame rate, CMOS sensors de nitely outperform CCDs.
Yet, there are still important sectors where CCDs lead the market, like imaging in the near infrared
(700 to 1000nm). In this case, imagers need to have a thicker photon absorption region. Most CMOS
imager fabrication processes are tuned for high volume applications that only image in the visible.
Increasing the substrate thickness will degrade the ability of the imager to resolve spatial features.
Conversely, CCDs can be fabricated with thicker absorbing layers while preserving their ability to
resolve ne spatial features. Hence, CCDs that are speci cally designed to be highly sensitive in the
near infrared are much more sensitive than CMOS imagers beyond 700 nm.
In the following we synthetically report a comparative analysis of CCD and CMOS image sensors
considering some of the most important features.
• Quantum e ciency (QE). On this point, CCD and CMOS imagers can be considered at the
same level, but CCD has historically bene ted from many years of technological process re ne-
ment for QE enhancement. CMOS imagers are generally made on technologies inspired from
those used for integrated circuits and generally do not use speci c QE optimisation strategies.
However, recently the implementation of speci c processes has brought to QE improvement in
CMOS sensors, leading to QE relatively close to that obtained with high-end CCD.
• Responsivity. CMOS imagers are marginally superior to CCDs, in general, because gain ele-
ments are easier to place on a CMOS image sensor. Their complementary transistors allow low-
power high-gain ampli ers, whereas CCD ampli cation usually comes at a signi cant power
penalty.
• Dynamic range. It gives CCDs an advantage over CMOS imagers when operated at low speed.
CCDs still enjoy lower noise because of quieter sensor substrates due to less on-chip circuitry
and common output ampli ers with transistor geometries that can be easily adapted for minimal
noise. However, when the sensors are operated at high frame rate, CMOS imagers outperform
CCDs because of a lower readout noise that stems from the intrinsic parallelism of the charge to
voltage conversion process.
• Uniformity. Spatial wafer processing variations, defects and ampli er variations create non-uni-
formities in the response of pixels under identical illumination conditions. CMOS imagers were

186
fi

fi

ffi
fi
fi
fi

fi
fi
fi

fi
fi
ff
fi
ff
fi
fi
fi
fi

fi
fi

traditionally worse that CCDs. Each pixel had an open-loop output ampli er, and the o set and
gain of each ampli er varied considerably because of wafer processing variations. However, feed-
back-based ampli ers can trade o gain for greater uniformity under illumination. Such ampli-
ers have made the uniformity of some CMOS imagers closer to that of CCDs. O set variation of
CMOS ampli ers, though, still manifests itself as non-uniformity in darkness, which is worse
than that of CCDs.
• Shuttering. The ability to start and stop exposure arbitrarily is a standard feature of virtually all
consumer and industrial CCDs, especially interline transfer devices, and is particularly important
in machine vision applications. CCDs can deliver superior electronic shuttering, with little ll-
factor compromise, even in small-pixel image sensors. Implementing uniform electronic shutter-
ing in CMOS imagers requires a number of transistors in each pixel. In area-scan imagers, uni-
form electronic shuttering comes at the expense of ll factor because the opaque shutter transis-
tors must be placed in what would otherwise be an optically sensitive area of each pixel. CMOS
matrix sensor designers have dealt with this challenge in two ways: A nonuniform shutter, called
a rolling shutter, exposes di erent lines of an array at di erent times. It reduces the number of
in-pixel transistors, improving ll factor. This is sometimes acceptable for consumer imaging, but
in higher-performance applications, object motion manifests as a distorted image (Figure 4.40).
A uniform synchronous shutter (global shutter) exposes all pixels of the array at the same time.
Object motion stops with no distortion, but this approach consumes pixel area because it re-
quires up to 5 transistors in each pixel .
• Speed. In this regard, CMOS arguably has a large advantage over CCDs because readout is mas-
sively parallel and all camera functions can be placed on the image sensor. Moreover, the native
digital output of CMOS imagers makes much easier to reach a high throughput compared to
CCDs, which generate analog signals.
• Windowing. One unique capability of CMOS technology is the ability to read out a portion of
the image sensor. This allows elevated frame or line rates for small regions of interest. This is an
enabling capability for CMOS imagers in some applications, such as high speed object tracking in
a subregion of the eld of view. CCDs, generally speaking, have limited windowing capability.
• Anti-blooming. The ability to drain localised overexposure without compromising the rest of
the image. CMOS generally has natural blooming immunity. CCDs, on the other hand, require
speci c engineering to achieve this capability. Many CCDs that have been developed for con-
sumer applications do, but those developed for scienti c applications generally do not.

Figure 4.40 Global shutter vs. rolling shutter in CMOS


based imagers.

187
fi
fi
fi
fi
fi
fi

ff
fi
ff

fi
fi
ff
fi
ff

ff

fi
• Biasing and clocking. CMOS imagers have a clear strength in this regard. They generally oper-
ate with a single bias voltage and clock level. Non-standard biases are generated on-chip, while
CCDs typically require a few higher voltage biases, but clocking has been simpli ed in modern
CCDs that operate with low-voltage clocks. CMOS imagers have lower power dissipation than
equivalent CCDs.
A summary of CCD and CMOS characteristics is presented in Table 1. Some features are to the
bene t of one or the other technology, without a real supremacy in overall performance. However, the
point de nitely di erentiating CMOS is the exibility of implementation with the system-on-chip
(SOC) approach and the lower power consumption; on the other hand, in high end applications,
where extremely low noise is required, CCDs are still the detectors of choice.
Finally, we want to spend few words to compare the readout noise features of CCDs and CMOS
imagers. The noise is the result of two contributions: 1/f icker noise (low frequency) and thermal
noise (high frequency), as shown Figure 4.41. The icker noise is the consequence of charge being
trapped in the gate oxide. The traps are present because of defects generated by the technological
process. The lling and emptying of these traps lead to uctuations in current owing in the transis-
tor channel. The CCD readout noise can be extremely low for detectors designed for applications
such as astronomy, for which the image is read at a low frequency. The system design includes elec-
tronics whose frequency bandwidth is minimised to avoid integration of the time uctuations of the
signal. For these applications the 1/f component of noise is dominant. For high speed video ap-
plications, noise is much higher due to the overwhelming contribution of the thermal component
that leads to a signi cant degradation of the signal to noise ratio. The CMOS image sensor has the
advantage of a column parallel readout scheme (see Figure 4.39). The readout frequency is therefore
divided by the number of columns compared with CCD. Consequently, the readout noise of CMOS
imagers is generally dominated by the 1/f contribution. Figure 4.42 shows the behaviour of the
readout noise as a function of the sampling frequency. Despite the progress in this area, the sensitivi-
ty of CMOS image sensors is still limited by the readout noise for extremely low light applications.
Conversely, for CCDs, the electron multiplication in EMCCD has brought to an e ective readout

Table 4.1 Summary of CCD and CMOS characteristics.

188
fi
fi
fi

ff
fi
fl
fl
fl
fl

fl
fi
fl
ff
noise of a fraction of electron r.m.s., which make this technology the best option for demanding sci-
enti c applications. In general applications, CCDs have been gradually replaced by CMOS imagers
and in the future EMCCD can potentially move to electron multiplying CMOS (EMCMOS). Like
EMCCD, this technology is planned to be used to improve image quality at the extremely low light
levels required for scienti c use

Figure 4.41 Power spectral density of the readout


noise.

Figure 4.42 Readout noise in CCD and CMOS image Sen-


sors (CIS) as a function of the sampling frequency.

189
fi
fi

You might also like