You are on page 1of 14

General method for simulating laboratory tests

with constitutive models for geomechanics


Tomáš Janda1 and David Mašı́n2
1
Czech Technical University in Prague, Faculty of Civil Engineering, Czech Republic
2
Charles University in Prague, Faculty of Science, Czech Republic

August 18, 2016

1 Abstract

2 The paper describes an approach to the simulation of arbitrary labo-


3 ratory tests with homogeneous stress and strain fields applicable to most
4 constitutive models common in geomechanics. The method by Bardet and
5 Choucair [Int. J. Numer. Anal. Meth. Geomech. 15(1):119, 1991] is gen-
6 eralized for an arbitrary number of controlling and controlled variables.
7 The approach is illustrated for time-dependent thermo-hydro-mechanical
8 constitutive models. The purpose of the method is to define an endless
9 variety of different laboratory tests declaratively by means of two matrices
10 E and S, which define the constraints on the controlled quantities such
11 as stress, strains, suction or temperature.

12 1 Introduction
13 Most of the constitutive models for mechanical, thermal or coupled analysis can
14 be written in the general rate form

σ̇ g = g σ̇ (ε̇g , σ g , q) (1)
q̇ = g q̇ (ε̇g , σ g , q) (2)

15 where εg denotes the controlling variables vector containing all independent


16 (prescribed) quantities and σ g is the controlled variables vector containing all
17 dependent (unknown) quantities. Vector q represents state variables and (˙) =

1
d()
18
dt denotes the time derivative. For time independent models, t is the pseudo-
19 time (integration variable).
20 For both time-dependent or time-independent thermo-hydro-mechanical con-
21 stitutive models the above quantities can be written as

σg = {σ11 , σ22 , σ33 , σ12 , σ13 , σ23 , Sr }T (3)


εg = {ε11 , ε22 , ε33 , ε12 , ε13 , ε23 , s, T }T (4)

22 where σij are components of the stress tensor (in the present hydro-mechanical
23 case σij represents a net stress, calculated as the total stress minus the pore air
24 pressure), Sr is the degree of saturation, εij are components of the strain tensor,
25 s is suction and T is temperature. In general, the controlling and controlled
26 variables vectors can have any number of components. Number of components
27 of the controlling variables tensor will be denoted as n (in Eq. (7), n = 8) and
28 number of components of the controlled variables tensor will be denoted as nσ
29 (in Eq. (7), nσ = 7).
30 It is to be pointed out that the vectors εg and σ g are not classical ”strain”
31 and ”stress” vectors which would have to be work-conjugated, instead they are
32 variables chosen purely based on convenience. As an example, let us consider
33 partially saturated soil at constant temperature. For such a soil, Houlsby [2]
34 derived the following equations for work-conjugated generalised stress and strain
35 tensors:

0 0 0 0 0 0
σg = {σ11 , σ22 , σ33 , σ12 , σ13 , σ23 , ns}T (5)
T
εg = {ε11 , ε22 , ε33 , ε12 , ε13 , ε23 , Sr } (6)

36 where n is porosity and σ is so-called Bishop stress calculated as

σ0 = σ tot − 1 (ua − Sr s) (7)

37 where σ tot is total stress, ua is pore air pressure and 1 is second-order identity
38 tensor.
39 Still, as discussed for example by Sheng et al. [6], in conventional finite ele-
40 ment codes the displacements, pore pressures and temperatures are first solved
41 at nodal points. The strains at integration points are then solved from the dis-
42 placements. Then, purely for the model implementation, it is most convenient
43 to consider strains, suction and temperature as controlling variables and stresses

2
44 and degree of saturation as controlled variables.
45 To obtain the model’s response to the prescribed increment of the controlling
46 variables we need to integrate the governing rate equations (1) and (2). The
47 simplest technique to achieve this is employing the forward Euler integration
48 scheme, but more advanced methods such as the adaptive Runge-Kutta methods
49 can also be utilized. For a finite increment of the controlling variables vactor
50 and an initial values of the controlled variables vector and state variables we
51 formally write the resulting increment of stress and state variables in the form
Z t2
∆σ g = g ∆σ (∆εg , σ g , q) = g σ̇ (ε̇g (t), σ g , q)dt (8)
t1
Z t2
∆q = g ∆q (∆εg , σ g , q) = g q̇ (ε̇g (t), σ g , q)dt (9)
t1

52 where each of the functions g ∆σ and g ∆q covers the integration of both equa-
53 tions (1) and (2) over a finite increment of the controlling variables vector ∆εg ,
54 t1 and t2 are times at the beginning and at the end of the current step. In the
55 case of rate-dependent models, t1 and t2 represent real times and g σ̇ and g q̇
56 are defined in terms of real rates, whereas in rate-independent materials time is
57 merely an integration variable.
58 The above formulation is sufficient for the simulation of laboratory tests in
59 which only the components of the controlling variables vector are prescribed.
60 When the laboratory test requires to prescribe or constrain also the components
61 of the controlled variables vector we exploit an approach proposed Bardet and
62 Couchair [1]. The change of the controlling variables vector, controlled variables
63 vector and state variables due to a single loading step has to satisfy n , possibly
64 nonlinear, equations

S∆σg + E∆εg = ∆y (10)

65 where E is a matrix of n rows and columns and S is a matrix of n rows and


66 nσ columns. The vector ∆y has n components. It has only the first of its
67 components non-zero, thus:

∆y = {∆y, 0, . . . , 0}T (11)

68 where ∆y controls the size of the load increment. The first row of matrix S
69 resp. E assign the load increment ∆y to one or more of the components of

3
70 vector ∆σg or ∆εg . The rest of the rows of S and E define the constraints
71 applied on vectors ∆σg or ∆εg during a particular test. Note that this technique
72 requires the laboratory test having only one degree of freedom, i.e. all quantities
73 controlled during the test are in fact expressed in terms of a single scalar quantity
74 ∆y. It will be shown later that most laboratory tests satisfy this criterion and,
75 if not, the vector ∆y can be made more general. Obviously, this technique is
76 valid only for laboratory tests in which we can assume homogeneous stress and
77 strain fields within the sample.

78 2 Nonlinear solution using Newton–Raphson


79 method
80 The function g ∆σ (∆εg , σ g , q) is nonlinear for most of the nontrivial constitutive
81 models. Therefore – for laboratory tests that prescribe the dependent variables
82 ∆σ g – the whole system of equations (10) becomes nonlinear and has to be
83 solved iteratively. A standard approach to the solution of such a system is the
84 Newton–Raphson method which searches for the root of vector-valued function
85 that in our case becomes

f (∆εg ) = Sg ∆σ (∆εg , σ g , q) + E∆εg − ∆y = 0 (12)

86 We solve the equations for a prescribed finite load increment ∆y.


87 From the implementation point of view, we need – apart form the ability to
88 evaluate the function f (∆εg ) – to follow three steps to solve the above equation
89 by means of the Newton–Raphson method:

90 • Provide a Jacobian matrix, i.e. matrix of first derivatives, of f (∆εg )

∂fi (∆εg )
Jij (∆εg ) = , (13)
∂∆εg,j

91 where fi (∆εg ) is the i-th component of the vector-valued function f (∆εg )


92 and ∆εg,j is the j-th component of its argument vector ∆εg . For models,
93 where function f (∆εg ) is not differentiable, such as the models based on
94 the theory of hypoplasticity, we need to provide an appropriate approxi-
95 mation of Jacobian matrix.

96 • Set the initial value of the strain increment ∆εg . It is reasonable to chose

4
97 the initial value ∆εj=0
g = 0.

98 • Formulate a condition upon which the iterative algorithm stops. The stop-
99 condition requires special attention. In principle, we can use the Euclidean
100 norm of f and stop the iterations when it is below a certain tolerance, i.e.

101 f · f ≤ T OL. This approach is, however, not recommended here since
102 the rows of equation (12) are the residua of potentially different quantities.
103 For such a setup it is difficult to come up with an appropriate tolerance
104 as it is applied jointly to stress components, strain components and other
105 quantities in vectors ∆εg and ∆σ g . We therefore prefer to formulate the
106 stop condition solely in terms of ∆εg and use the form

e · e ≤ T OL, (14)

107 where e is a vector of the normalized residua defined as

e = RJ −1 f (∆εg ), (15)

108 with R = diag(1, . . . , 1, 1/rs , 1/rT ) being a diagonal matrix which merely
109 normalizes the last two non-strain components corresponding to suction
110 and temperature by factors rs and rT , respectively. It should be noted
111 that only the error of expression RJ −1 f is checked to be small instead of
112 the original requirement of f to be small. This should be considered in
113 cases where the components of J are large, such as in very stiff materials.

114 The algorithm for the Newton–Raphson method is summarized as

5
i = 0;
εig = εinit
g ;
σ ig = σ init
g ;
q i = q init ;
r i = 0;
// Loop over loading steps
while i < numLoadSteps do
J −1 = J −1 (σ i , q i );
f (∆εg ) = f (∆εg , σ ig , q i ) + r i ;
j = 0;
∆εjg = 0;
// Loop over Newton–Raphson iterations

115
while e · e > T OL do
∆εj+1
g = ∆εjg − J −1 f (∆εjg );
j = j + 1;
end
r i+1 j
∆σ = f (∆εg )
σ i+1
g = σ ig + f ∆σ (∆εjg , σ ig , q i );
q i+1 = q i + f ∆q (∆εjg , σ ig , q i );
εi+1
g = εig + ∆εjg ;
i=i+1
end
Algorithm 1: Forward integration of the constitutive model. Index i corre-
sponds to a loading step, index j denotes the iteration of the Newton–Raphson
method.

116 Note that after the stop condition is satisfied, in the case of controlled stress
117 components, there still remain residua r of the function f which differ from
118 the required zero values. Accuracy of the algorithm is significantly improved
119 (by several orders of magnitude), if these residua increments are considered as
120 targets in the subsequent load step. How they are introduced is clear from the
121 algorithm above. The introduction of residua into the function f (∆εg ) prevent
122 the inaccuracies of the strain increment ∆εg to be accumulated along multiple
123 loading steps as illustrated in Fig. 1.
124 Also note that the corrections using residua, as suggested and implemented
125 in this paper, is only necessary when the Newton–Raphson algorithm targets
126 load increments without reflecting the global level of loading. Checking the
127 globally defined quantities is more standard approach, which is usually em-

6
128 ployed within the nonlinear finite element method. This more conventional
129 approach can also be utilized in the presented context but it would require
130 more complicated setup for load paths in which the load controlled values are
131 not proportional, for example the stress paths with a constant direction starting
132 from the non-zero stress state.

Figure 1: Schema of the Newton–Raphson iterations in the first load step re-
sulting in residua r1 (left) and the iterations in i-th step with function f shifted
by ri (right).

133 3 Examples of S and E matrices


134 In this section, for the sake of illustration, we define the matrices S and E
135 for two basic laboratory tests, namely the oedometric test and drained triaxial
136 shear test. Matrices S and E for various other laboratory tests are defined in
137 Appendix.

138 3.1 Oedometric test


139 For a strain controlled oedometer test with a time independent material the
140 matrices read
8×7
S = 0 , (16)
8×8
E = 1 , (17)

141 where 0 is a matrix with zero elements and 1 is the identity matrix. This way the
142 load increment ∆y corresponds to the prescribed increment of strain εxx while
143 the remaining components of ∆εg are kept zero. The system of equations does

7
144 not impose any other constrains to components of the controlled variables vector
145 ∆σ g . In this special type of laboratory test we control only the independent
146 variables ∆εg and the dependent variables ∆σ g are just calculated by the model
147 via equation (8).

148 3.2 Drained triaxial test


149 The shearing phase of the strain controlled drained triaxial test is characterized
150 by an increasing vertical strain while the lateral and shear components of stress
151 are kept constant. The nonzero elements of the matrices E and S read

E11 = E77 = E88 = 1, (18)


S22 = S33 = S44 = S55 = S66 = 1. (19)

152 With these matrices the first equation of the system (10) prescribes the load
153 increment ∆y to the first component of the controlling variables vector εg , i.e.
154 the normal strain component εxx . All components of stress apart from the
155 vertical one are prescribed as constants together with the last two components
156 of the controlling variables vector ∆εg , i.e. the suction s and temperature T .

157 4 Demonstration of the algorithm performance


158 The proposed approach will be demonstrated by simulating three different ex-
159 periments using an advanced thermo-hydro-mechanical double structure hy-
160 poplastic model for expansive clays. The model, proposed by Mašı́n [5, 4], is an
161 evolution of the hydro-mechanical double structure model proposed in Ref. [3].
162 This model has been selected as it is highly non-linear, and due to its complex
163 nature only a crude estimate of the Jacobian J is available. To be more specific,
164 in the present implementation, mechanical part of the Jacobian has been given
165 by 3fs L, where fs is hypoplastic scalar factor and L is hypoplastic fourth-order
166 tensor (more in [5, 4]). The multiplier 3 has been selected by a trial-and-error
167 procedure to make sure that non-linearity induced by the non-linear compo-
168 nents of the hypoplastic model does not imply the actual stiffness to be higher
169 then the Jacobian. The non-mechanical components of the Jacobian have been
170 found by direct perturbation.
171 Figure 2 shows the results of one of the experiments simulated (heating-
172 cooling cyclic test at constant net stress and suction), compared with experi-

8
exp., s=110 MPa
exp., s=39 MPa
exp., s=9 MPa
model, s=110 MPa
model, s=39 MPa
model, s=9 MPa
80

70

Temperature [°C]
60

50

40

30
p=0.1 MPa
20
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
εv [%]

Figure 2: Simulations of heating-cooling test at constant mean net stress p and


suction s using thermo-hydro-mechanical double structure hypoplastic model,
compared with experimental data on MX80 bentonite (figure from [5, 4], data
from [7]).

173 mental data by Tang et al. [7]. Details of the model and parameters adopted
174 are in [5, 4].
175 The proposed algorithm has been implemented into an in-house element test
176 driver, which adopts the forward Euler scheme for the model integration. The
177 following three experiments have been simulated, which all have at least one
178 controlled variables vector component prescribed, so Newton–Raphson itera-
179 tions are needed to solve Equation (12).

180 • Drained triaxial test: axisymmetric test with controlled axial strain incre-
181 ment. Horizontal stresses, suction and temperature are held constant.

182 • Water retention wetting test: Suction-controlled experiment with constant


183 net stress (equal to 0) and temperature.

184 • Heating-cooling cyclic test: Temperature-controlled experiment with con-


185 stant net stress and suction.

186 The simulations have been run at sufficiently small forward Euler step size
187 (such that the step size has minor effect on predictions). Figures 3 to 5 show
188 the dependency of the controlled stress component (a) and number of iterations
189 (b) on the controlling variables vector components for three values of the tol-
190 erance T OL. Clearly, even relatively high tolerance T OL=10−2 leads to minor
191 fluctuations in the prescribed stress component, and practically perfect match

9
100.5 35
TOL=1e-4
100.4 30 TOL=1e-3
TOL=1e-2

Number of iterations [-]


100.3
25
100.2
σr [kPa]

100.1 20
100 15
99.9
10
99.8
TOL=1e-4
99.7 TOL=1e-3 5
TOL=1e-2
99.6 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3
εa [-] εa [-]

(a) (b)

Figure 3: Simulations of drained triaxial test at constant suction and temper-


ature, dependency on tolerance T OL. (a) evolution of horizontal stresses, (b)
number of iterations.

0.04 25
TOL=1e-4 TOL=1e-4
0.035 TOL=1e-3 TOL=1e-3
TOL=1e-2 TOL=1e-2
20
Number of iterations [-]

0.03

0.025
pnet [kPa]

15
0.02

0.015 10

0.01
5
0.005

0 0
0 20000 40000 60000 80000 100000 120000 0 20000 40000 60000 80000 100000 120000
suction [kPa] suction [kPa]

(a) (b)

Figure 4: Simulations of water retention wetting test at zero net stress and
constant temperature, dependency on tolerance T OL. (a) evolution of mean
net stress, (b) number of iterations.

192 is reached for lower tolerance of T OL=10−4 . This good performance is tightly
193 linked to the consideration of residual vector r in the algorithm: two to three
194 orders of magnitude lower values of T OL would be needed to reach similar
195 level of accuracy without consideration of r. Due to the crude estimation of
196 the Jacobian J , a relatively high number of iterations is required to reach the
197 specified tolerance. As expected, the number of iterations increases with a de-
198 creasing tolerance T OL. The simulations have been run with rs = 10000 and
199 rT = 100, but it is noted that in the present case these values have no effect on
200 the predictions.

10
100.004 250
TOL=1e-4 TOL=1e-4
100.003 TOL=1e-3 TOL=1e-3
TOL=1e-2 TOL=1e-2
200

Number of iterations [-]


100.002
100.001
pnet [kPa]

150
100
99.999
100
99.998
99.997 50
99.996
99.995 0
290 300 310 320 330 340 350 360 290 300 310 320 330 340 350 360
T [K] T [K]

(a) (b)

Figure 5: Simulations of heating-cooling test at constant net stress and suction,


dependency on tolerance T OL. (a) evolution of mean net stress, (b) number of
iterations.

201 5 Conclusions
202 In this paper, a general algorithm was introduced for predicting various labo-
203 ratory experiments using element-test driver. Using the method, it is possible
204 to define an endless variety of different laboratory tests declaratively by the
205 means of two matrices E and S. Numerical solution of the system of equations,
206 adopting the Newton–Raphson integration scheme and considering the effect of
207 residual vector r, has also been introduced. The scheme yields accurate results
208 for quite large tolerance values T OL, even for highly non-linear models adopting
209 an inaccurate estimation of the Jacobian J .

210 Acknowledgment
211 The authors are grateful for the financial support by the grant TACR TA04031603.
212 The second author also acknowledges support by the grant GACR 15-05935S.

213 References
214 [1] J. P. Bardet and W. Choucair. A linearized integration technique for in-
215 cremental constitutive equations. Int. J. Numer. Anal. Meth. Geomech.,
216 15(1):1–19, 1991.

217 [2] G. T. Houlsby. The work input to an unsaturated granular material.


218 Géotechnique, 47(1):193–196, 1997.

11
219 [3] D. Mašı́n. Double structure hydromechanical coupling formalism and a
220 model for unsaturated expansive clays. Engineering Geology, 165:73–88,
221 2013.

222 [4] D. Mašı́n. Coupled thermo-hydro-mechanical double structure model for


223 expansive soils. Journal of Engineering Mechanics (under review), 2016.

224 [5] D. Mašı́n. Development of a coupled thermo-hydro-mechanical double struc-


225 ture model for expansive soils. In Delage, P. et al., editor, 3rd European
226 Conference on Unsaturated Soils, Paris, France (in print), 2016.

227 [6] Sloan S. Sheng, D. and A. Gens. A constitutive model for unsaturated soils:
228 thermomechanical and computational aspects. Computational Mechanics,
229 6:453–465, 2004.

230 [7] A.-M. Tang, Y.-J. Cui, and N. Barnel. Thermo-mechanical behaviour of a
231 compacted swelling clay. Géotechnique, 58(1):45–54, 2008.

232 Appendix
233 In this Appendix matrices E and S are defined for various laboratory tests. In
234 the following, we define non-zero components of the matrices E and S in the
235 index notation (row number, column number, counted from 1) (all the other
236 components are zero).

237 • Axial strain controlled undrained triaxial shear stress (constant volume).

E11 = E31 = E32 = E33 = E77 = E88 =1 (20)


S22 = S44 = S55 = S66 =1 (21)
S23 = −1 (22)

238 • Stress-controlled isotropic compression test.

E77 = E88 =1 (23)


S11 = S21 = S31 = S44 = S55 = S66 =1 (24)
S22 = S33 = −1 (25)

12
239 • Axial strain controlled constant deviatoric stress test at constant suction
240 and temperature.

E11 = E77 = E88 =1 (26)


S21 = S31 = S44 = S55 = S66 =1 (27)
S22 = S33 = −1 (28)

241 • Axial strain controlled constant mean stress test at constant suction and
242 temperature.

E11 = E77 = E88 =1 (29)


S21 = S22 = S23 = S32 = S44 = S55 = S66 =1 (30)
S33 = −1 (31)

243 • ε12 controlled drained (constant vertical stress) simple shear test at con-
244 stant suction and temperature.

E14 = E22 = E33 = E55 = E66 = E77 = E88 =1 (32)


S41 = =1 (33)

245 • ε12 controlled undrained (constant volume) simple shear test at constant
246 suction and temperature.

E14 = E22 = E33 = E41 = E55 = E66 = E77 = E88 =1 (34)

247 • Suction-controlled constant net stress and temperature test.

E17 = E88 =1 (35)


S22 = S33 = S44 = S55 = S66 = S71 =1 (36)

248 • Temperature-controlled constant net stress and suction test.

E18 = E77 =1 (37)


S22 = S33 = S44 = S55 = S66 = S81 =1 (38)

249 • Suction-controlled swelling pressure (constant volume) test at constant

13
250 temperature.

E17 = E22 = E33 = E44 = E55 = E66 = E71 = E88 =1 (39)

251 • Suction-controlled constant vertical stress oedometric test at constant


252 temperature (swelling strain test).

E17 = E22 = E33 = E44 = E55 = E66 = E88 =1 (40)


S71 =1 (41)

253 • Temperature-controlled heating pressure (constant volume) test at con-


254 stant suction.

E18 = E22 = E33 = E44 = E55 = E66 = E77 = E81 =1 (42)

14

You might also like