Research Article
doi.org/10.1002/prep.202200001
Machine-Learning a Solution for Reactive Atomistic
Simulations of Energetic Materials
Rebecca K. Lindsey,*[a] Cong Huy Pham,[a] Nir Goldman,[a, b] Sorin Bastea,[a] and Laurence E. Fried[a]
Abstract: Many of the safety and performance-related
properties of energetic materials (EM) are related to com-
plex condensed phase chemistry at extreme P,T conditions
eluding direct experimental investigation. Atomistic simu-
lations can play a vital role in generating insight into EM
chemistry, but they rely critically on the availability of suit-
able interatomic potentials (“force fields”). The ChIMES ma-
chine learning approach enables generation of interatomic
potentials for condensed phase reacting systems, with ac-
curacy similar to Kohn-Sham density functional theory
through its unique, highly flexible orthogonal basis set of
interaction functions and systematically improvable many-
Keywords: atomistic simulations · machine learning · reactive force field · interatomic model · ChIMES · chemistry
1 Introduction
Atomistic simulations have proven to be a critical tool for
energetic materials (EM) research over the past several dec-
ades, since they can provide insight into processes occur-
ring under the extreme and rapidly changing thermody-
namic conditions associated with EM function, where direct
experimental probing is often difficult or impossible [1].
This simulation approach has been applied to a wide varie-
ty of application areas including predicting material equa-
tion of state (EOS) [2], determining nanoscale phenomena
governing material evolution at the grain scale [3], and pro-
viding insight into the mechanisms controlling EM safety,
such as the formation of hot spots under shock loading [4–
6]. Ultimately, atomistic simulations could accelerate efforts
toward developing new energetic materials with improved
safety/energy trade-offs [7].
Nevertheless, atomistic simulations present their own
challenges. EM evolution under extreme conditions is large-
ly governed by chemical reactivity occurring on the nano-
meter and picosecond spatiotemporal scales, yet character-
istic scales for relevant emergent phenomena (e. g., defect
formation, pore collapse, and carbon condensation) can
span up to four orders of magnitude [8]. Reconciling the in-
herently multiscale nature of EM research has traditionally
required atomistic simulators to choose between highly
predictive yet computationally intensive first principles (FP)
approaches like Kohn-Sham [9] density functional theory
(i. e., which confine accessible scales to a few hundred
atoms and 10s of picoseconds), or efficient molecular me-
Propellants Explos. Pyrotech. 2022, 47, e202200001 (1 of 11)
pep.wiley-vch.de
body expansion of interatomic interactions. ChIMES has
been successfully applied to a variety of systems including
simple model energetic materials, both as a correction for
simpler quantum theory and as a stand-alone interatomic
potential. In this perspective, the successes and challenges
of applying the ChIMES approach to the reactive molecular
dynamics of energetic materials are outlined. Our machine-
learned approach is general and can be applied to a variety
of different application areas where atomic-level calcu-
lations can be used to help guide and elucidate experi-
ments.
chanics (MM) simulations using interatomic potential (IAPs)
[10, 11] capable of modeling EM on spatiotemporal scales
approaching 1 μm and 1 μs. Non-reactive MM IAPs, such as
those developed by Bedrov et al. [12] are highly efficient
and accurate for non-reactive processes such as thermal
transport in solid energetics [13] while reactive MM IAPs
(e. g., the widely used ReaxFF [14]) have been applied to a
variety of energetic molecules and processes [4, 10, 11].
Although strides in the scientific communities’ under-
standing of EM evolution under extreme conditions have
been made using these approaches, increased interest in
high accuracy modeling of reactive processes in EM on larg-
er spatiotemporal scales demands more of atomistic simu-
lation capabilities. For example, there is a need to rapidly
evaluate new and/or candidate materials via highly accu-
rate and efficient crystal structure searches [15], understand
hot spot initiation mechanisms on experimental scales [16],
and elucidate carbon condensation mechanisms in non-ide-
al EM [17]. All of these applications require an agile model
development capability; however, standard MM IAP devel-
[a] R. K. Lindsey, C. Huy Pham, N. Goldman, S. Bastea, L. E. Fried
Physical and Life Sciences Directorate
Lawrence Livermore National Laboratory
Livermore, California 94550, USA
*e-mail: rklindsey@llnl.gov
[b] N. Goldman
Department of Chemical Engineering
University of California, Davis
Davis, California 95616, USA
© 2022 Wiley-VCH GmbH
Machine-Learning a Solution for Reactive Atomistic Simulations of Energetic Materials
opment approaches tend to be slow and yield models of
either limited chemical transferability or accuracy. Ulti-
mately, IAPs capable of bridging the gap between accuracy
and efficiency of FP and MM methods are needed before
these problems can be gainfully pursued.
Machine learning (ML), or more specifically, machine-
learned interatomic potentials (ML–IAPs) provide a promis-
ing path forward. Machine learning encompasses a broad
class of data analytic approaches wherein the fitting frame-
work “learns” to construct models that make predictions
based on some set of data descriptors, with the minimal
human intervention [18]. Within the context of IAPs, the
“data descriptors” generally summarize the atomic coor-
dinates, (e. g., pair distances, bending angles, and embed-
ding densities in the case of MM-IAPs), and are used to pre-
dict system energy, atomic forces, and system stress.
Quantum simulations can be used to generate the training
data set, and parameters can be calibrated via, e. g., the
“force matching” approach [19]. Traditional MM–IAPs relate
a given set of atomic coordinates to the system potential
energy through a fixed mathematical expression. Typically,
this expression has a characteristic potential energy func-
tion shape and a fixed number of parameters. In contrast,
ML–IAPs use a collection of basis functions to represent the
potential energy and contain many more parameters (e. g.,
typically in the range 103–105) than is typical for MM–IAPs,
making them profoundly more flexible. For example, a sin-
gle ML–IAP framework can be used to describe non-re-
active, reactive, metallic, and/or insulating systems [20–23],
whereas MM–IAPs generally require a different set of equa-
tions to describe each different system class. Note that this
flexibility is particularly helpful in cases where the appro-
priate function form isn’t known a priori, which is partic-
ularly relevant for the extreme conditions produced by
functioning energetic materials. ML–IAPs are generally ca-
pable of arbitrary complexity and high accuracy, although
electronic degrees of freedom such as spin are typically ne-
glected. In addition, ML–IAPs can be developed as correc-
tions to approximate quantum methods such as DFTB [24–
28] or to MM–IAPs such as ReaxFF [29]. Typically, correction
approaches require fewer parameters and simpler func-
tional forms than ML–IAPs that completely describe the IAP.
The advent of ML–IAPs has had a transformative effect
on broader atomistic simulation research and has stimu-
lated the development of diverse underlying forms and
methods [20, 29, 30]. Artificial neural network (ANN) ap-
proaches [31] have experienced a surge of popularity due
to widely available open-source toolkits. However, the bulk
of model development and application efforts have focused
on inorganic/solid-state materials or materials under mild
conditions relative to those associated with EM function.
EM ML–IAP development presents several unique chal-
lenges due to the need to describe condensed phase re-
activity in chemically complex organic systems. In partic-
ular, shocked materials experience a broad range of
thermodynamic conditions, varying from ambient temper-
Propellants Explos. Pyrotech. 2022, 47, e202200001 (2 of 11)
ature to 1000 s of K and 10,000s of MPa, which can drive
the system from insulating to metallic [32], induce chemis-
try, and result in phase separation [33, 34]. At the same
time, energetic materials can include four or more ele-
ments, which further complicates chemistry, necessitates
determining 103–105 of model parameters [35], and compli-
cates generation of a suitable training dataset. Ultimately,
developing ML–IAPs for EM is a difficult task. In this per-
spective, practical challenges in EM ML–IAP development
are presented along with an overview of possible solutions
and outcomes, and discussion of remaining challenges and
next steps, within the context of the ChIMES approach.
2 Methods
2.1 ChIMES Overview: Functional Form
The goal in developing ChIMES is to establish a practical
and tractable alternative to traditional IAPs that enables:
* Efficient and accurate simulations of novel HE
* Minimized time to determine optimal model parameters
* Maximized simulation efficiency
* Description of complicated reactive phenomena
The ChIMES formalism draws upon insights from
chemistry that bonding can be described in terms of a rela-
tively small number of interacting atoms. This suggests that
an expansion in the number of interacting atoms could
yield a functional form that is simpler, faster to parameter-
ize, and more computationally efficient than common ANN
approaches.
The design philosophy behind ChIMES involves map-
ping quantum mechanical energies onto linear combina-
tions of many-body Chebyshev polynomials of the first
kind. These polynomials have a number of desirable proper-
ties for creating interatomic potential energy surfaces, in-
cluding: (i) they are orthogonal with respect to a weighting
function and can be generated recursively, allowing for ba-
sis set completeness and user-defined complexity, (ii) high-
er-order polynomials tend to have decreasing expansion
coefficient values (due to their monic form), and (iii) they
are “nearly optimal” in that the error in an expansion will
closely resemble a minimax polynomial. In addition, de-
rivatives of these polynomials are related to Chebyshev pol-
ynomials of the second kind, which themselves are orthog-
onal, and can also be generated recursively. This allows for
easy and reliable determination of forces and stress tensor
components for atomistic calculations.
Briefly, the ChIMES total energy corresponds to an n-
body expansion:
© 2022 Wiley-VCH GmbH
Research Article R. K. Lindsey, C. Huy Pham, N. Goldman, S. Bastea, L. E. Fried

X
na X
na X
na the f c smoothly varying cutoff functions for each con-
E nB ¼ 1
E i1 þ 2
E i1 i2 þ 3
E i1 i2 i3 þ � � � stituent pair distance. Penalty functions are not included in
i1 i1 >i2 i1 >i2 >i3 this case and instead are handled entirely by the two-body
X
na (1) interaction. Note that additional details on extending
nB
þ Ei1 i2 ���in B
ChIMES to greater than three-body interactions can be
i1 >i2 ��� inB 1 >inB found in [57].
Optimal ChIMES parameters (i. e., polynomial co-
efficients) are generated through nominal “force matching”
where EnB is the total ChIMES system energy, nB is the max- to per-atom force components, overall energies, and overall
imum bodiedness, n Ei1 i2 :::in is the n-body ChIMES energy for a stress tensors for configurations from DFT-MD simulations.
given set of atoms with indices i ¼ fi1 ; i2 ; :::; in g, and na is Since ChIMES models are parametrically linear, this is ac-
the total number of atoms in the system. The one-body en- complished by solving the overdetermined matrix equation
ergies, 1 Ei1 , correspond to the atomic energy constants for AC ¼ B, where the matrix A corresponds to the values of
each element type. the requisite polynomials for a given training configuration
The two-body (pairwise) energies are expressed as linear (i. e., derivatives with respect to the fitting coefficients). The
combinations of Chebyshev polynomials of the first kind: column vectors C and B correspond to the linear ChIMES
coefficients and the training forces, energies, and stresses,
� �X
O2B respectively. Note that this linear least-squares optimization
2 e e �
Ei1 i2 ¼ f p ri1 i2 þ f eci ei ri1 i2 Cemi ei T m si1i i2 i
1 2 1 2 1 2
(2) problem can be solved with any number of algorithms [28].
m¼1 The ChIMES cluster-based description thus provides a
chemically-motivated means of explicit many-body descrip-
e e �
In this case, T m si1i i2 i represents a Chebyshev poly- 1 2
tion while the Chebyshev polynomials afford a flexible, con-
e e
nomial of order m, and si1i i2 i is the pair distance transformed 1 2
vergent orthogonal basis with several mathematically favor-
to exist over the interval ½ 1; 1� using � a Morse-like function able features. Moreover, ChIMES models are parametrically
e e
[36]. Here, si1i i2 i / exp ri1 i2 =le1 e2 and le1 e2 is an element-
1 2
linear, hence advanced linear solvers (e. g., which perform
pair distance scaling constant, usually taken to be the peak parameter selection) can be used to rapidly generate mod-
position of the first coordination shell. Ceni ei is the corre- 1 2
els. In addition to use as a stand-alone ML–IAP, note that
sponding permutationally invariant coefficient for the inter- ChIMES can be used to generate corrections to other IAPs,
action between atom types ei1 and ei2 taken from�the set of e. g., as a Density Functional Tight Binding correction en-
all possible element types, feg. The term f eci ei ri1 i2 is a Ters- 1 2
abling application to materials under extreme conditions.
off cutoff function [37] which forces 2 Ei1 i2 to zero beyond a
maximum distance defined for a given fe1 ; e2 g pair type. To
prevent sampling of ri1 i2 distances below � the model inner 2.2 ChIMES Overview: Use as a DFTB Correction
cutoff, a smooth penalty function f p ri1 i2 is introduced. The
reader is referred to previous work for additional details The Density Functional Tight Binding method (DFTB) [38] is
[22]. a semi-empirical quantum simulation approach that can be
Greater than two-body orthogonal basis sets are gen- viewed as an efficient proxy for full-quantum calculations
erated via products of the n choose 2 unique constituent for both gas- [24] and condensed phase systems [39–41].
pairwise polynomials of the higher-order terms. For exam- The computational efficiency of DFTB (i. e., 102–103 improve-
ple, a three-body term has 3 choose 2 ¼ 3 pairs, which ment over DFT) arises from its approximate Hamiltonian
yields the following expression for the ChIMES three-body and use of (generally) short-ranged empirical functions,
energy: though it still requires solving for electronic eigenstates,
� � � which generally scales as OðN3 Þ, similar to DFT calculations.
3
Ei1 i2 i3 ¼ f eci ei ri1 i2 f eci ei ri1 i3 f eci ei ri2 i3 �
1 2 1 3 2 3
Nevertheless, the computational efficiency of DFTB has
0 a two-fold immediate advantage: (1) simulations can be run
X
O3B X
O3B X
O3B
� � � (3)
C ei1 ei2 ei3
Tm s
ei1 ei2
Tp s
ei1 ei3
Tq s
ei2 ei3 for several orders of magnitude longer than what is feasible
mpq i 1 i2 i1 i3 i2 i 3
m¼0 p¼0 q¼0 with standard DFT-MD simulations (e. g., nanoseconds vs.
tens of picoseconds, e. g., relevant to high-pressure carbon
The above expression contains a triple sum for the condensation [42–44]), and (2) large numbers of statistically
i1 i2 ; i1 i3 ; i2 i3 polynomials over the hypercube up to order independent reactive MD simulations can be run simulta-
O3B , and include a single permutationally invariant co- neously, allowing for the generation of ensemble statistics.
efficient for each set of powers and atom types, Cempq i ei ei
. The 1 2 3
DFTB begins by assuming spherically symmetric charge
primed sum denotes that only terms for which two or more densities centered on atoms and then expanding the Kohn-
of the m; p; q polynomial powers are greater than zero are Sham DFT energy expression to second order in charge
included in order to guarantee that three distinct atom-cen- fluctuations [45]. Hamiltonian and overlap matrix elements
ters are evaluated. The expression for 3 Ei1 i2 i3 also contains are pre-tabulated from a two-center interaction potential

Propellants Explos. Pyrotech. 2022, 47, e202200001 (3 of 11) © 2022 Wiley-VCH GmbH
Machine-Learning a Solution for Reactive Atomistic Simulations of Energetic Materials
with use of a minimal local atomic basis set. The remaining
double counting and ion-ion repulsion terms are then
grouped into a “repulsive energy”, which is generally fit to
short-range empirical functional forms (usually spanning
bonded distances or those of the first solvation shell, only).
The resulting energy expression can be written as follows:
EDFTB ¼ EBS þ ECoul þ ERep þ EDisp :
Here, EBS is the band structure energy, determined from
the occupied electronic eigenstates, ECoul is the charge
transfer term, and ERep is the previously described repulsive
energy. EDisp is an externally added dispersion energy that
can be used to improve accuracy for weakly interacting ma-
terials [46]. The empirical nature of ERep allows DFTB to be
“tuned” to specific physical and/or chemical properties such
as isothermal compression curves or small cluster data, fre-
quently computed from DFT or even higher-level quantum
chemical data.
DFTB models generally exist within a “sweet-spot” be-
tween DFT and completely empirical MD models in terms
of accuracy and ease of development. The quantum me-
chanical part of the calculation tends to be relatively accu-
rate for many conditions and materials, allowing for a rea-
sonable representation of the system’s electronic ground
state [28]. This is especially true under extreme conditions,
where electron repulsions tend to dominate interactions
and exchange-correlation interactions do not need to be
determined as precisely [47]. In addition, the empirical ERep
a function is only fit to very short-ranged interactions,
which yields several advantages. Due to the underlying
quantum mechanics, DFTB models generally require far less
training data than full ML–IAPs, and generally exhibit a high
degree of transferability across materials and thermody-
namic conditions [48]. For example, a recent model for bulk
α-Ti as well as bulk TiH2 and its surface chemistry was de-
termined from several dozen MD snapshots performed on
the TiH2 12 atom unit cell over a range of pressures [28]. In
this regard, DFTB has potential as an intermediate model-
ing approach, where some initial investment in determining
an accurate ERep can result in a highly efficient model that
retains most of the accuracy of higher levels of theory.
These models can be used as a proxy for full DFT calcu-
lations for both determinations of structural and chemical
information, as well as to generate significantly larger quan-
tities of training data to ultimately be used for the determi-
nation of inter-atomic potentials [26].
Consequently, the most significant difficulty with creat-
ing DFTB models frequently resides with a determination of
ERep . In the past, ERep has been determined based on a clus-
ter or high-symmetry crystalline data (e. g., cubic lattice
compression curves) to pair-wise functional forms, such as a
simple polynomial power series or cubic spline [39]. How-
ever, these approaches are unable to correct for the greater
than two-center terms that can be significant for some sys-
tems [49]. This is particularly true for materials under ex-
Propellants Explos. Pyrotech. 2022, 47, e202200001 (4 of 11)
treme conditions, where sudden pressurization and heating
can cause transient metallization and different bonding ar-
rangements can lead to significant changes in electronic
states and the resulting chemical reactivity [50].
In this respect, ChIMES presents a convenient tool for in-
cluding additional many-body effects within ERep . This can
allow models to overcome deficiencies due to the reduced
(4) dimensionality of the DFTB approximate Hamiltonian and
minimal basis set by matching features such as atomic
forces, system energies, and stress tensor terms. Briefly,
ChIMES can be used for ERep determination by first creating
a DFTB repulsive energy-specific training set, where DFTB
forces and stress tensor components with the repulsive en-
ergy set to zero are subtracted from their respective DFT
values, e. g.:
* ¼ Ft
FtRep FtQM;DFTBa
a DFTa
i i i
(5)
* ¼ st
stRep aa DFTaa stQM;DFTBaa : (6)
* ¼ Et
EtRep EtQM;DFTBaa : (7)
aa DFTaa
Here, t corresponds to a specific MD configuration, a to
the cartesian direction, and i is the atomic index. The ‘*’ is
used to denote that the quantities being computed are part
of the training set, and ‘QM,DFTB’ refers to the quantum
components of the DFTB calculation, i. e., only forces, en-
ergies, and stresses from EBS and ECoul :
ChIMES parameters can then be determined from mini-
mizing the following objective function:
vffiffiffiffiffiffi0
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffihffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiiffiffiffiffi1
ffiffiffiffi
u PM PN P3 2
u t t*
u B t¼1 i¼1 a¼1 FChIMESa FRepa C
u B i i
C
u B h i2 C
u1 B P C
Fobj ¼u þ Mt¼1 EtChIMES EtRep * C; (8)
uN B B C
u B d
C
u @ h i A
t PM P3 2
t t*
þ s ChIMESaa s Repaa
t¼1 a¼1
where M is the total number of configurations in the train-
ing set, and Nd is the total number of data entries (3MN
force components plus 3M stress tensor components).
ChIMES/DFTB models have been created for several sys-
tems, including high-level quantum chemical organic clus-
ters [27], detonating energetic materials [26], and metallic
systems and their surfaces [25, 28].
2.3 ChIMES Overview: Special Considerations
Several practical considerations factor into the develop-
ment of ML–IAPs. Perhaps most importantly, one must con-
sider whether the provided basis and training data are
compatible with key governing physics in the target sys-
tem. For example, the standard ChIMES form (e. g., Eq’s 1–3
above) combined with Newtonian molecular dynamics do
© 2022 Wiley-VCH GmbH
Research Article
not explicitly account for spin, long-range charge inter-
actions, quantum nuclear vibrational effects, or variable
electronic temperature. In many cases, these assumptions
are reasonable (e. g., in the cases of tightly coupled spin
state and ion configuration, and charge screening effects);
however, in cases where they are not, supplemental terms
should be added [22, 51, 52].
Other considerations include the need to strike a bal-
ance between accuracy and efficiency. In principle, increas-
ing model complexity enables improved recovery of train-
ing data (overfitting notwithstanding); however, doing so
concurrently increases computational expense. Identifying
that balance is a non-trivial endeavor frequently performed
using methods like holdout cross-validation, but this can be
difficult when trying to select from 103–105 parameters. To
overcome this challenge, advanced linear solvers are em-
ployed, (e. g., LARS and LASSO) [53–55] which use absolute
value-based regularization to automatically set parameters
deemed inconsequential to zero; this has the effect of de-
creasing model size while increasing both stability and
computational efficiency. Note that these algorithms can be
distributed for application to large fitting problems (e. g.,
104–105 parameters) [23].
2.4 ChIMES Overview: On Accuracy, Efficiency,
Robustness, and Agility
As with any ML–IAP framework, ChIMES model develop-
ment accuracy, efficiency, robustness, and agility is largely
governed by three interrelated factors:
* Model complexity
* Model optimization
* Training data generation efficiency
Broadly speaking, increasing model complexity (i. e.,
number of fitting parameters) enables improved model ac-
curacy; however, large models are computationally in-
efficient and generally challenging to optimize due to the
expansive parameter space (this is especially true for non-
linear models). This is further complicated by the fact that
ML models “learn” these parameters solely from the avail-
able training data. Irrespective of model complexity, an in-
adequately large and/or informative training set will pre-
clude developing an accurate (and possibly, physically
reasonable) model. Ultimately, training data plays a critical
role in determining ML–IAP accuracy and robustness; how-
ever, its generation remains a grand challenge, raising
questions including:
* How can training data be generated in an efficient man-
ner?
* How can maximally informative training data be identi-
fied?
In section 2.2 above, use of ChIMES-corrected DFTB as
an efficient proxy for DFT during training data generation
and/or labeling was discussed. Below, additional solutions
Propellants Explos. Pyrotech. 2022, 47, e202200001 (5 of 11)
R. K. Lindsey, C. Huy Pham, N. Goldman, S. Bastea, L. E. Fried
for agile training data generation that have been in-
corporated into the ChIMES framework are described.
ChIMES training data is typically generated via DFT-MD.
Models trained to these data will excel at describing favor-
able regions of physicochemical space but can struggle to
describe relatively unfavorable regions (i. e., corresponding
to relatively rare events), which are likely to be under-repre-
sented in the training set. For some materials and applica-
tions (e. g., non-reactive systems, solid-state materials, and
gas-phase chemistry), supplementary training config-
urations can be generated via straightforward means (e. g.,
well-established enhanced sampling methods, random dis-
placements, and manual perturbations to system lattice pa-
rameters). However, generating suitable configurations for
energetic materials is substantially more challenging since
models must describe chemically reactive condensed phase
molecular systems characterized by disparate energy scales
(i. e., governing bonded, non-bonded, and reactive inter-
actions) and an expansive configurational space for the
many molecular species formed.
For relatively simple materials, iterative refinement can
be used [22, 56] to generate the target configurations. In
this approach, an initial model is fit to an initial training set
(Figure 1, parts 1 and 2), then used to launch simulations
(Figure 1, part 3). If the resulting physical predictions are in-
sufficiently accurate, a set of configurations from the result-
ing trajectories can then be combined with the initial train-
ing set, once DFT forces, energies, and stresses are assigned
(Figure 1, parts 4 and 6). The updated training set can then
be used to train a new model (Figure 1 parts 1 and 2), and
the cycle is repeated until the desired model accuracy is
achieved. Note that ChIMES is particularly well suited to this
approach since models can be generated rapidly. For more
complex fitting problems (e. g., higher chemical complex-
ity), more advanced approaches are needed. In these cases,
active learning can greatly accelerate model generation, by
selecting specific configurations and/or molecular struc-
tures most likely to be underrepresented in the training set
(Figure 1. part 5) [57].
3 Results and Discussion
A central goal of ChIMES is enabling an agile and robust
framework for developing quantum-accurate models for EM
capable of predicting material thermochemistry, the shock
Hugoniot equation of state, and shockwave phenomena in-
cluding carbon coagulation and hot spot initiation. The ap-
plication of ChIMES to systematically more complex materi-
als and problems has simultaneously enabled evaluation of
the underlying framework efficacy and an improved under-
standing of the chemistry driving detonation-induced ma-
terial response.
In this section, an overview of some of these ChIMES
modeling efforts is provided, including:
* Understanding shockwave-driven carbon condensation
© 2022 Wiley-VCH GmbH
Machine-Learning a Solution for Reactive Atomistic Simulations of Energetic Materials

Figure 1. The ChIMES active learning framework.

* Sensitization in liquid explosives oxygen purification kinetics) [33, 34]. Moreover, these simu-
* QM-accurate gas-phase chemistry calculations lations were validated through UF experiments, which veri-
fied the predicted cluster sizes, shape, chemical composi-
tion, and formation time scales (see Figure 2) [33]. This
3.1 Understanding Shockwave-Driven Carbon effort has provided a powerful validation of ChIMES and ap-
Condensation plication to chemically complicated problem spaces. This
effort has also served as a starting point to systematically
3.1.1 Shocked Carbon Monoxide investigate more complex problems, including carbon con-
densation in the hydrogen-free EM 3,4-bis(3-nitrofurazan-4-
Non-ideal EM can release up to 10 % of their detonation en- yl)furoxan (DNTF), discussed in greater detail below.
ergy through carbon condensation, which can have sig-
nificant implications for EM performance and sensitivity.
However, this complicated condensed-phase reaction-driv- 3.1.2 Shocked DNTF
en process remains poorly understood due to the asso-
ciated spatiotemporal scales (i. e., spanning 0.1 to 10s of ChIMES is also enabling improved understanding of EM for
nanometers and up to 100 s of nanoseconds), which have which experimental data are limited. For example, a
so far, precluded direct experimental observation. Though
considered rapid by experimental standards, these scales
are coincidentally at the upper limit of accessibility to atom-
istic simulation methods, requiring quantum accurate (de-
fined here as having an accuracy similar to Kohn-Sham DFT)
models capable of simulation on large scale. This process is
further complicated by high chemical complexity (i. e., the
manifold of chemistry that unravels as C/H/O/N-containing
materials begin to react under high T and P). To better un-
derstand this process, a reductionist investigation of carbon
condensation in shock compressed cryogenic liquid carbon
monoxide was conducted, through a hybrid experimental
and simulation study. In particular, ChIMES models [56] ca-
pable of simulation on the same time scales and over-
lapping length scales as those achievable through ultra-
fast-laser-driven shock (“UF”) experiments were developed.
These models were used to predict shock-induced Figure 2. Left: ChIMES simulations predicting nanocarbon for-
mation on experimentally relevant scales. Black and red beads are
nano-carbon formation over 10s of ps timescales, and pro-
carbon and oxygen atoms, respectively; atom connections are
vided an in-depth description of the overall system evolu- drawn as a visual aid and do not necessarily correspond to chem-
tion (e. g., cluster size distribution, composition, and classi- ical bonds. Right: TEM image from UF experiments. Red semicircles
cal governing growth kinetics) associated chemical kinetics indicate individual nanocarbon particles, for which a simulation-
(e. g., reactive transport of carbon into and between clus- predicted image is also shown in the zoomed-in dashed yellow out-
ters, the evolution of the fluid surrounding clusters, and line.

Propellants Explos. Pyrotech. 2022, 47, e202200001 (6 of 11) © 2022 Wiley-VCH GmbH
Research Article
ChIMES-corrected DFTB model was developed for DNTF
that was transferable to T/p states relevant to detonation,
with the aim of shedding light on questions including: (1)
what are the major chemical kinetic steps associated with
its shock decomposition? (2) what is its Hugoniot equation
of state? (3) how do these properties compare with those
for other carbon-rich EM? and (4) what are the implications
for nanocarbon formation mechanisms in DNTF and other
EM? For this study, a ChIMES-corrected DFTB model (i. e.,
rather than a full-ChIMES model) was developed for en-
hanced agility, since the underlying semi-empirical quan-
tum method significantly reduces the ChIMES model com-
plexity (and correspondingly, training set size) requirements
to generate a high accuracy model.
To address these questions, the ChIMES-corrected DFTB
model was leveraged in extended Lagrangian Multiscale
Shock Technique (MSST) simulations [58–60], which con-
strain the sampled thermodynamic states to those for a
steady planar shockwave within continuum theory (i. e.,
simulating steady shock waves by constraining the stress
and energy of an MD simulation to the Rayleigh line and
the Hugoniot energy relations). This simulation approach
enabled estimation of the DFT Chapman-Jouguet (C J) det-
onation state, Hugoniot curve, and shockwave-driven
chemistry [26]. On DFTB-accessible scales, DNTF chemical
evolution was found to be largely characterized by release
of CO, CO2, and N2, and in contrast to oxygen-rich and oxy-
gen-balanced materials [42, 61], NO3 does not form in sig-
nificant quantities nor appear to inhibit chemical evolution.
Formation and slow evolution of large CNO-containing spe-
cies (i. e., which are likely precursors to the experimentally
observed carbon condensates) were also found to be a
characteristic feature of shock-driven DNTF chemistry. Ulti-
mately, this combined ChIMES/DFTB/MSST simulation ap-
proach provides a means of reaching far longer timescales
with “quantum accuracy.” Beyond accelerating first-princi-
ples-based EM simulation studies, this approach can also
significantly expedite development efforts for stand-alone
ChIMES ML–IAP models, by serving as an efficient, effective
means of generating DFT-quality training and validation
data for materials where experimental data are scarce.
3.2 Sensitization in Liquid Explosives
More recently, ChIMES models have been created to study
shocked liquid hydrazoic acid (HN3, i. e., the simplest co-
valent azide-based EM) for which previous DFTB MSST sim-
ulations suggested rapid detonation chemistry [50]. In a
previous study [23], a ChIMES model capable of accurately
reproducing DFT-predicted material structure, dynamical
properties, and chemistry for a wide range of unreactive
and decomposition conditions of liquid HN3 was developed.
This model was then leveraged in MSST simulations to
study shockwave-driven detonation of liquid HN3. ChIMES
predicted a C J detonation velocity between 6.5 and
Propellants Explos. Pyrotech. 2022, 47, e202200001 (7 of 11)
R. K. Lindsey, C. Huy Pham, N. Goldman, S. Bastea, L. E. Fried
6.6 km/s, i. e., closer to the experimentally reported range of
7.1–7.6 km/s than the value predicted using standard DFTB
models (6.0 km/s) [50], demonstrating that ChIMES can out-
perform the accuracy of semi-empirical quantum ap-
proaches. Moreover, the model is approximately 5 orders of
magnitude more computationally efficient than DFT and
scales linearly with system size, enabling simulations for
larger spatiotemporal scales, as was done in the presently
described 1 ns MSST simulations of a 2048 atom system.
Here, preliminary results for direct-shock simulations of
porous liquid HN3 using ChIMES are presented. Though
more computationally demanding than the previous MSST
simulations due to use of over 50-fold more atoms, it is a
more powerful method for direct observation of phenom-
ena like hotspot formation. LAMMPS [62] with the ChIMES
calculator library [63] was used to simulate an orthogonal
computational cell with initial parameters a = 16.52 Å, b =
155.18 Å, and c = 622.77 Å and a total of 108,800 atoms
(27,200 molecules). Periodic boundary conditions are ap-
plied in the x- and y-axis while in the z-axis where a shock-
wave is generated, a non-periodic boundary condition is
used. A fixed wall is placed at the left edge of the simu-
lation box in the z-axis and a cylindrical pore with the radi-
us r = 40 Å is created 200 Å from the fixed wall. An initial
velocity -up is provided for all atoms in the NVE simulation.
The particles impact the wall and reflect, generating a
shockwave that propagates along the z-axis.
Figure 3 shows the evolution of system density as a
shockwave passes through the material for the simulation
with up = 1.5 km/s. After about 7.5 ps, the pore has col-
lapsed, and hotspot formation begins. As can be seen in
Figure 3, density at hot spot location exceeds that of both
the shocked and unshocked material. As the hot spot re-
laxes, rarefaction waves are observed, momentarily creating
a region near the fixed wall that has lower density than that
of the shocked regions. The Shock Trapping Internal Boun-
daries (STIBs) method [6] was used to study the evolution of
the hot spot, in which small immobile material regions are
imposed between the shockwave and the hot spot. These
immobile STIBs regions have the widths of 20 Å, which is
wider than the cutoff distance for force interactions be-
tween particles (8 Å in this case); atoms outside of these re-
gions (i. e., the hot spot and its surrounding material is in
between the STIBs) are fully mobile. With these constraints,
atoms of interest interact with each other and with the im-
mobile STIB atoms, allowing for direct study of the time
evolution of the hot spot. Material outside two STIBs re-
gions are neglected to reduce computational cost. Note
that the STIBs method must be applied when all shock-
waves are sufficiently far from the area of interest, which in
this case, it is at the simulation time of 10 ps.
The evolution of the system temperature in the STIBs
simulation for porous HN3 is shown in the right panel of
Figure 4. For comparison, a similar simulation for liquid HN3
was performed with no initial void, and the result is also
displayed in Figure 4 (left panel). For porous HN3, the pore
© 2022 Wiley-VCH GmbH
Machine-Learning a Solution for Reactive Atomistic Simulations of Energetic Materials
Figure 3. Time evolution (i. e., from top to bottom in 2.5 ps incre-
ments) of a 108,800 atom HN3 system containing a cylindrical void
shocked via impact with a piston to the left of the system. System
density is indicated in g cm 3 by colors in the color bar. At 10 ps,
the STIBs method was applied to the area of interest (i. e., contain-
ing the hot spot), between the two vertical red lines shown in the
bottom panel.
Figure 4. Time evolution (i. e., from top to bottom in 25 ps incre-
ments) of the temperature in the STIBs simulations for pure liquid
HN3 (left panel) and porous HN3 (right panel). System temperature
is indicated in K by colors in the color bar.
collapse creates a hot spot with the temperature in some
regions > 2000 K, while the temperature of the shocked
material is only ~ 750 K. At later times, the hot spot diffuses
its heat to surrounding atoms and the system temperature
becomes uniform after 50 ps. This is in contrast with the
Propellants Explos. Pyrotech. 2022, 47, e202200001 (8 of 11)
simulation of the liquid HN3 (without pore collapse) where
the system temperature is nearly uniform over the whole
material. In both cases, the system temperature increases in
time as chemical reactions occur.
These preliminary results indicate that the presence of a
pore significantly affects the evolution of the shocked sys-
tem. Overall, this study shows the ability of ChIMES simu-
lations to provide quantum accurate results for shocked
materials that would not be possible using quantum me-
chanical calculation directly.
3.3 Quantum-Accurate Gas-Phase Chemistry Calculations
Recently, a DFTB/ChIMES model was developed for organic
materials, by training to a small subset (i. e., only ~ 0.3 %) of
the ANI-1x dataset, which contains a diverse set of molec-
ular conformation at equilibrium and non-equilibrium geo-
metries, with forces and energies assigned via hybrid DFT
[35]. As described in section 2.2, the benefit of using
ChIMES in this manner is that many-body interactions can
be corrected for in a systematic and rapidly tunable process
by varying the ChIMES polynomial order parameters. Addi-
tionally, the underlying semi-empirical framework enables
the determination of the electronic states of a system, al-
lowing for possible thermal electronic excitations, which
can be significant for many energetic materials under shock
compression and can lower activation energies for re-
activity.
The performance of the DFTB/ChIMES model in predict-
ing energies, forces and vibrational frequencies has also
been tested on many additional quantum datasets. Ulti-
mately, this model yielded strong agreement with reference
data from either hybrid DFT, coupled-cluster calculations or
experiments. While all training data for developing DFTB/
ChIMES were taken from gas-phase clusters only, this model
has also been shown to reproduce the experimental struc-
ture of graphite (a well-known challenge for DFT).
Here, new results for the enthalpy of formation (ΔHf) of
EM (Figure 5) predicted by standard DFTB (with the 3ob pa-
rameter set) and a DFTB/ChIMES model developed in [64],
using protocols described in [27, 64] are presented. As
shown in [64], the highly accurate ccCA-PS3 method yields
excellent agreement with available experimental data, with
a mean absolute error (MAE) of ~ 12,500 J/mol. DFTB pre-
dicts ΔHf in good agreement with experimental data for
some cases (for example RDX and HMX). DFTB/ChIMES
shows an improvement over DFTB in all cases considered
here. The MAE values for DFTB and DFTB/ChIMES are
~ 42,000 and 25,000 J/mol, respectively. For the molecules
where experimental data is not available (NTO, DADE and
LLM-105), DFTB/ChIMES predicts the values closer to that of
ccCA-PS3 than DFTB. It is worth mentioning that the accu-
racy of the DFTB/ChIMES model is very similar to G3 and G4
quantum composite methods used in [64], although it uses
only a fraction of computational resources required by G3
© 2022 Wiley-VCH GmbH
Research Article
Figure 5. Comparison of predicted enthalpies of formation of some
energetic molecules for DFTB, DFTB/ChIMES, ccCA-PS3 (ref. 64) with
experimental data (ref. 67) and predictions from an atom-equiv-
alent method based on quantum chemical calculations at the
B3LYP/6-31G* level (Rice et al., ref. 68).
or G4. For example, PETN quantum composite calculations,
which involve geometry optimization and vibrational fre-
quency generation for 100s of trial configurations, G3 and
G4 requires approximately 500 GB of physical memory dis-
tributed among 32 CPUs, for 21 days. In contrast, DFTB/
ChIMES calculations including geometry optimization and
frequency determination use a single node and take about
30 minutes.
4 Conclusion
There is clearly a need for highly accurate and scalable
atomistic simulations of EM spanning temperatures from
~ 300 to ~ 5000 K and pressures from 0 to ~ 50,000 MPa.
Across this range of conditions, there are tremendous
changes in speciation and chemical reactivity. The ChIMES
ML–IAP provides a promising path forward and has been
successfully applied to a number of systems including mod-
el energetic materials. Within the ChIMES family, DFTB/
ChIMES ML–IAPs can be generated most rapidly and require
minimal training sets, while the stand-alone ChIMES ML–
IAPs are scalable to millions of atoms. Further development
of ChIMES is required to access the full range of species and
states typical of reacting CHNO energetic materials, such as
RDX or TATB, which is the subject of ongoing work.
A key challenge common to all ML–IAP development ef-
forts is that the DFT-MD simulations typically used to gen-
erate training data over-sample statistically favorable con-
figurations while unfavorable configurations important for
chemistry, such as transition states, are under-sampled.
Data selection through the ChIMES active learning frame-
Propellants Explos. Pyrotech. 2022, 47, e202200001 (9 of 11)
R. K. Lindsey, C. Huy Pham, N. Goldman, S. Bastea, L. E. Fried
work is helpful in this regard, but additional development is
needed for full generalization to arbitrary materials.
A second challenge is that achieving quantum-level ac-
curacy for highly complex chemistry may require the addi-
tion of explicit � 5-body interactions in the potential energy
function, whereas ChIMES applications to date have in-
cluded a maximum of 4-body terms. This is a straightfor-
ward generalization of the models presented here but
would represent a significant increase in the number of pa-
rameters. Handling a very large number of possible poten-
tial parameters will require very scalable linear regularized
optimization and parameter selection implementations
(such as LASSO), and highly effective methods to generate
training configurations.
Finally, the ChIMES approach does not currently treat
some quantum aspects of material behavior such as elec-
tron spin and electron temperature. Nuclear quantum vibra-
tional effects could be treated through path-integral molec-
ular dynamics [65] or quantum thermostats [66], but have
not yet been investigated with the ChIMES framework.
Acknowledgements
This work was performed under the auspices of the U.S. Depart-
ment of Energy by Lawrence Livermore National Laboratory under
Contract No. DE-AC52-07NA27344. LLNL-JRNL-830464.
Data Availability Statement
The data that support reported findings are available within
all cited references and/or the manuscript body.
References
[1] M. R. Manaa, L. E. Fried, The reactivity of energetic materials
under high pressure and temperature, in Advances in Quantum
Chemistry (Ed. J. Sabin), Academic Press, London 2014, p. 221.
[2] A. C. Landerville, M. W. Conroy, M. M. Budzevich, Y. Lin, C. T.
White, I. I. Oleynik, Equations of state for energetic materials
from density functional theory with van der Waals, thermal,
and zero-point energy corrections. Appl. Phys. Lett. 2010, 97,
251908.
[3] R. Menikoff, T. D. Sewell, Constituent properties of HMX need-
ed for mesoscale simulations. Combust. Theory Model. 2002, 6,
103.
[4] Q. An, S. V. Zybin, W. A. Goddard, A. Jaramillo-Botero, M. Blan-
co, S. N. Luo, Elucidation of the dynamics for hot-spot ini-
tiation at nonuniform interfaces of highly shocked materials.
Phys. Rev. B – Condens. Matter Mater. Phys. 2011, 84, 220101.
[5] T. R. Shan, R. R. Wixom, A. P. Thompson, Extended asymmetric
hot region formation due to shockwave interactions following
void collapse in shocked high explosive. Phys. Rev. B 2016, 94,
054308.
© 2022 Wiley-VCH GmbH
Machine-Learning a Solution for Reactive Atomistic Simulations of Energetic Materials
[6] B. W. Hamilton, M. P. Kroonblawd, C. Li, A. Strachan, A Hot-
spot’s Better Half: Non-Equilibrium Intra-Molecular Strain in
Shock Physics. J. Phys. Chem. Lett. 2021, 12, 2756.
[7] J. M. Winey, Y. Toyoda, Y. M. Gupta, Near-optimal combination
of high performance and insensitivity in a shock compressed
high explosive single crystal. J. Appl. Phys. 2021, 130, 015902.
[8] B. C. Barnes, J. K. Brennan, E. F. C. Byrd, S. Izvekov, J. P. Lar-
entzos, B. M. Rice, Toward a Predictive Hierarchical Multiscale
Modeling Approach for Energetic Materials, in Challenges and
Advances in Computational Chemistry and Physics (Ed. J. Leszc-
zynski), Springer, Cham, 2019 p. 229.
[9] W. Kohn, L. J. Sham, Self-consistent equations including ex-
change and correlation effects. Phys. Rev. 1965 140, A1133.
[10] A. Strachan, E. M. Kober, A. C. T. Van Duin, J. Oxgaard, W. A.
Goddard, Thermal decomposition of RDX from reactive molec-
ular dynamics. J. Chem. Phys. 2005, 122, 054502.
[11] D. Guo, S. V. Zybin, A. An, W. A. Goddard, F. Huang, Prediction
of the Chapman-Jouguet chemical equilibrium state in a deto-
nation wave from first principles based reactive molecular dy-
namics. Phys. Chem. Chem. Phys. 2016, 18, 2015.
[12] D. Bedrov, G. D. Smith, T. D. Sewell, Temperature-dependent
shear viscosity coefficient of octahydro-1,3,5,7-tetranitro-
1,3,5,7-tetrazocine (HMX): A molecular dynamics simulation
study. J. Chem. Phys. 2000, 112, 7203.
[13] M. P. Kroonblawd, and T. D. Sewell, Theoretical determination
of anisotropic thermal conductivity for crystalline 1,3,5-tri-
amino-2,4,6-trinitrobenzene (TATB). J. Chem. Phys. 2013 139,
074503.
[14] A. C. T. Van Duin, S. Dasgupta, F. Lorant, W. A. Goddard, Re-
axFF: A Reactive Force Field for Hydrocarbons. J. Phys. Chem. A
2001, 105, 9396.
[15] B. A. Steele, S. M. Clarke, M. P. Kroonblawd, M. P., I. F. W. Kuo,
P. F. Pagoria, S. N. Tkachev, J. S. Smith, S. Bastea, L. E. Fried,
J. M. Zaug, E. Stavrou, O. Tschauner, Pressure-induced phase
transition in 1,3,5-triamino-2,4,6-trinitrobenzene (TATB). Appl.
Phys. Lett. 2019, 114, 191901.
[16] W. P. Bassett, B. P. Johnson, N. K. Neelakantan, K. S. Suslick,
D. D. Dlott, Shock initiation of explosives: High temperature
hot spots explained. Appl. Phys. Lett. 2017, 111, 061902.
[17] E. B. Watkins, K. A. Velizhanin, D. M. Dattelbaum, R. L. Gus-
tavsen, T. D. Aslam, D. W. Podlesak, R. C. Huber, M. A. Firestone,
B. S. Ringstrand, T. M. Willey, M. Bagge-Hansen, R. Hodgin, L.
Lauderbach, T. Van Buuren, N. Sinclair, P. A. Rigg, S. Seifert, T.
Gog, Evolution of Carbon Clusters in the Detonation Products
of the Triaminotrinitrobenzene (TATB)-Based Explosive PBX
9502. J. Phys. Chem. C 2017, 121, 23129.
[18] F. Emmert-Streib, Z. Yang, H. Feng, S. Tripathi, M. Dehmer, An
Introductory Review of Deep Learning for Prediction Models
With Big Data. Front. Artif. Intell. 2020, 3, p. 4.
[19] F. Ercolesi, J. B. Adams, Interatomic potentials from first-princi-
ples calculations: The force-matching method. EPL 1994, 26,
583.
[20] V. L. Deringer, A. P. Bartók, N. Bernstein, D. M. Wilkins, M. Cer-
iotti, G. Csányi, Gaussian Process Regression for Materials and
Molecules. Chem. Rev. 2021, 121, 10073.
[21] R. K. Lindsey, L. E. Fried, N. Goldman, ChIMES: A Force Matched
Potential with Explicit Three-Body Interactions for Molten Car-
bon. J. Chem. Theory Comput. 2017, 13, 6222.
[22] R. K. Lindsey, L. E. Fried, N. Goldman, Application of the
ChIMES Force Field to Nonreactive Molecular Systems: Water
at Ambient Conditions. J. Chem. Theory Comput. 2019, 15, 436.
[23] C. H. Pham, R. K. Lindsey, L. E. Fried, N. Goldman, Calculation of
the detonation state of HN3 with quantum accuracy. J. Chem.
Phys. 2020, 153, 224102.
Propellants Explos. Pyrotech. 2022, 47, e202200001 (10 of 11)
[24] M. Stöhr, L. Medrano Sandonas, A. Tkatchenko, Accurate Many-
Body Repulsive Potentials for Density-Functional Tight Binding
from Deep Tensor Neural Networks. J. Phys. Chem. Lett. 2020,
11, 6835.
[25] N. Goldman, B. Aradi, R. K. Lindsey, L. E. Fried, Development of
a Multicenter Density Functional Tight Binding Model for Plu-
tonium Surface Hydriding. J. Chem. Theory Comput. 2018, 14,
2652.
[26] R. K. Lindsey, S. Bastea, N. Goldman, L. E. Fried, Investigating
3,4-bis(3-nitrofurazan-4-yl)furoxan detonation with a rapidly
tuned density functional tight binding model. J. Chem. Phys.
2021, 154, 164115.
[27] C. Huy Pham, R. K. Lindsey, L. E. Fried, N. Goldman, High Accu-
racy Semi-Empirical Quantum Models Based on a Reduced
Training Set. ChemRXIV, 2021, doi: 10.26434/chemrxiv-2021-
2clcb.
[28] N. Goldman, K. E. Kweon, B. Sadigh, T. W. Heo, R. K. Lindsey,
C. H. Pham, L. E. Fried, B. Aradi, K. Holliday, J. R. Jeffries, B. C.
Wood, Semi-Automated Creation of Density Functional Tight
Binding Models through Leveraging Chebyshev Polynomial-
Based Force Fields. J. Chem. Theory Comput. 2021, 17, 4435.
[29] P. Yoo, M. Sakano, S. Desai, M. M. Islam, P. Liao, A. Strachan, A
Neural network reactive force field for C, H, N, and O systems.
npj Comput. Mater. 2021, 7, 1.
[30] M. A. Cusentino, M. A. Wood, A. P. Thompson, Explicit Multiele-
ment Extension of the Spectral Neighbor Analysis Potential for
Chemically Complex Systems. J. Phys. Chem. A 2020, 124, 5456.
[31] L. Zhang, J. Han, H. Wang, R. Car, E. Weinan, Deep Potential
Molecular Dynamics: A Scalable Model with the Accuracy of
Quantum Mechanics. Phys. Rev. Lett. 2018, 120, 143001.
[32] E. J. Reed, M. R. Manaa, L. E. Fried, K. R. Glaesemann, J. D. Joan-
nopoulos, A transient semimetallic layer in detonating nitro-
methane. Nat. Phys. 2008, 4, 72.
[33] M. R. Armstrong, R. K. Lindsey, N. Goldman, M. H. Nielsen, E.
Stavrou, L. E. Fried, J. M. Zaug, S. Bastea, Ultrafast shock syn-
thesis of nanocarbon from a liquid precursor. Nat. Commun.
2020 111, 1.
[34] R. K. Lindsey, N. Goldman, L. E. Fried, S. Bastea, Chemistry-
Mediated Ostwald Ripening in Carbon-Rich C/O Systems at Ex-
treme Conditions, Nat. Commun. 2022, 13, 1424.
[35] J. S. Smith, O. Isayev, A. E. Roitberg, ANI-1: an extensible neural
network potential with DFT accuracy at force field computa-
tional cost. Chem. Sci. 2017, 8, 3192.
[36] W. Wang, B. C. Shepler, B. J. Braams, J. M. Bowman, Full-dimen-
sional, ab initio potential energy and dipole moment surfaces
for water. J. Chem. Phys. 2009, 131, 054511.
[37] J. Tersoff, Empirical interatomic potential for carbon, with ap-
plications to amorphous carbon. Phys. Rev. Lett. 1988, 61, 2879.
[38] B. Hourahine, B. Aradi, V. Blum, F. Bonafé, A. Buccheri, C. Ca-
macho, C. Cevallos, M. Y. Deshaye, T. Dumitric, A. Dominguez,
S. Ehlert, M. Elstner, T. Van Der Heide, J. Hermann, S. Irle, J. J.
Kranz, C. Köhler, Y. Kowalczyk, T. Kubař, I. S. Lee, V. Lutsker, R. J.
Maurer, S. K. Min, I. Mitchell, C. Negre, T. A. Niehaus, A. M. N.
Niklasson, A. J. Page, A. Pecchia, G. Penazzi, M. P. Persson, J.
Řezáč, C. G. Sánchez, M. Sternberg, M. Stöhr, F. Stuckenberg, A.
Tkatchenko, V. W. Z. Yu, T. Frauenheim, DFTB + , a software
package for efficient approximate density functional theory
based atomistic simulations. J. Chem. Phys. 2020, 152, 124101.
[39] A. K. Akshay, E. Wadbro, C. Köhler, P. Mitev, P. Broqvist, J.
Kullgren, CCS: A software framework to generate two-body
potentials using Curvature Constrained Splines. Comput. Phys.
Commun. 2021, 258, 107602.
© 2022 Wiley-VCH GmbH
Research Article
[40] M. P. Lourenço, E. C. Dos Santos, L. G. M. Pettersson, H. A.
Duarte, Accurate SCC-DFTB Parametrization for Bulk Water. J.
Chem. Theory Comput. 2020, 16, 1768.
[41] G. Dolgonos, B. Aradi, N. H. Moreira, T. Frauenheim, An im-
proved self-consistent-charge density-functional tight-binding
(SCC-DFTB) set of parameters for simulation of bulk and mo-
lecular systems involving titanium. J. Chem. Theory Comput.
2010, 6, 266.
[42] M. R. Manaa, E. J. Reed, L. E. Fried, N. Goldman, Nitrogen-Rich
Heterocycles as Reactivity Retardants in Shocked Insensitive
Explosives. J. Am. Chem. Soc. 2009, 131, 5483.
[43] M. P. Kroonblawd, R. K. Lindsey, N. Goldman, Synthesis of func-
tionalized nitrogen-containing polycyclic aromatic hydro-
carbons and other prebiotic compounds in impacting glycine
solutions. Chem. Sci. 2019, 10, 6091.
[44] C. B. Cannella, N. Goldman, Carbyne Fiber Synthesis within
Evaporating Metallic Liquid Carbon. J. Phys. Chem. C 2015, 119
(37), 21605.
[45] M. Elstner, D. Porezag, G. Jungnickel, J. Elsner, M. Haugk, T.
Frauenheim, Self-consistent-charge density-functional tight-
binding method for simulations of complex materials proper-
ties. Phys. Rev. B - Condens. Matter Mater. Phys. 1998, 58, 7260.
[46] E. Caldeweyher, S. Ehlert, A. Hansen, H. Neugebauer, S. Spich-
er, C. Bannwarth, S. Grimme, A generally applicable atomic-
charge dependent London dispersion correction. J. Chem.
Phys. 2019, 150, 154122.
[47] N. Goldman, E. J. Reed, I. F. W. Kuo, L. E. Fried, C. J. Mundy, A.
Curioni, Ab initio simulation of the equation of state and ki-
netics of shocked water. J. Chem. Phys. 2009, 130, 124517.
[48] N. Goldman, L. E. Fried, L. Koziol, Using Force-Matched Poten-
tials To Improve the Accuracy of Density Functional Tight
Binding for Reactive Conditions. J. Chem. Theory Comput. 2015,
11, 4530.
[49] N. Goldman, S. Goverapet Srinivasan, S. Hamel, L. E. Fried, M.
Gaus, M. Elstner, Determination of a density functional tight
binding model with an extended basis set and three-body re-
pulsion for carbon under extreme pressures and temperatures.
J. Phys. Chem. C 2013, 117, 7885.
[50] E. J. Reed, A. W. Rodriguez, M. R. Manaa, L. E. Fried, C. M. Tarv-
er, Ultrafast detonation of hydrazoic acid (HN3). Phys. Rev. Lett.
2012, 109, 038301.
[51] L. Koziol, L. E. Fried, N. Goldman, Using force matching to de-
termine reactive force fields for water under extreme thermo-
dynamic conditions. J. Chem. Theory Comput. 2017, 13, 135.
[52] O. T. Unke, S. Chmiela, M. Gastegger, K. T. Schütt, H. E. Sauce-
da, K.-R. Müller, SpookyNet: Learning Force Fields with Elec-
tronic Degrees of Freedom and Nonlocal Effects. Nat. Commun.
2021, 12, 7273.
[53] R. Tibshirani, Regression Shrinkage and Selection Via the Lasso.
J. R. Stat. Soc. Ser. B 1996, 58, 267.
Propellants Explos. Pyrotech. 2022, 47, e202200001 (11 of 11)
R. K. Lindsey, C. Huy Pham, N. Goldman, S. Bastea, L. E. Fried
[54] B. Efron, T. Hastie, I. Johnstone, R. Tibshirani, H. Ishwaran, K.
Knight, J. M. Loubes, P. Massart, D. Madigan, G. Ridgeway, S.
Rosset, J. I. Zhu, R. A. Stine, B. A. Turlach, S. Weisberg, I. John-
stone, R. Tibshirani, Least angle regression. Ann. Stat. 2004, 32,
407.
[55] J. Friedman, T. Hastie, R. Tibshirani, Regularization paths for
generalized linear models via coordinate descent. J. Stat.
Softw. 2010, 33, 1.
[56] R. K. Lindsey, N. Goldman, L. E. Fried, S. Bastea, Many-body re-
active force field development for carbon condensation in C/O
systems under extreme conditions. J. Chem. Phys. 2020, 153,
054103.
[57] R. K. Lindsey, L. E. Fried, N. Goldman, S. Bastea, Active learning
for robust, high-complexity reactive atomistic simulations. J.
Chem. Phys. 2020, 153, 134117.
[58] E. J. Reed, L. E. Fried, J. D. Joannopoulos, A Method for Tract-
able Dynamical Studies of Single and Double Shock Com-
pression. Phys. Rev. Lett. 2003, 90, 4.
[59] E. J. Reed, L. E. Fried, M. R. Manaa, J. D. Joannopoulos, A Multi-
Scale Approach to Molecular Dynamics Simulations of Shock
Waves, in Chemistry at Extreme Conditions (ed. Riad Manaa, M.),
Elsevier, 2005, p. 297.
[60] E. J. Reed, L. E. Fried, W. D. Henshaw, C. M. Tarver, Analysis of
simulation technique for steady shock waves in materials with
analytical equations of state. Phys. Rev. E – Stat. Nonlinear, Soft
Matter Phys. 2006, 74, 056706.
[61] N. Goldman, S. Bastea, Nitrogen oxides as a chemistry trap in
detonating oxygen-rich materials. J. Phys. Chem. A 2014, 118,
2897.
[62] S. Plimpton (1995) Fast parallel algorithms for short-range mo-
lecular dynamics. J. Comput. Phys. 117, 1, 1–19.
[63] rk-lindsey/chimes_calculator: Tools to interface ChIMES with
various external codes. https://github.com/rk-lindsey/chimes_
calculator.
[64] M. R. Manaa, L. E. Fried, I. F. W. Kuo, Determination of en-
thalpies of formation of energetic molecules with composite
quantum chemical methods. Chem. Phys. Lett. 2016, 648, 31.
[65] D. Marx, and M. Parrinello, Ab initio path integral molecular
dynamics: Basic ideas. J. Chem. Phys. 1996, 104, 4077.
[66] T. E. Markland, M. Ceriotti, Nuclear quantum effects enter the
mainstream. Nat. Rev. Chem. 2018, 2, 1.
[67] V. I. Pepkin, Y. N. Matyushin, Y. A. Lebdev, BACCA 1974, 23,
1707.
[68] B. M. Rice, S. V. Pai, J. Hare, Predicting heats of formation of
energetic materials using quantum mechanical calculations.
Combust. Flame 1999, 118, 445.
Manuscript received: January 1, 2022
Revised manuscript received: February 16, 2022
Version of record online: March 22, 2022
© 2022 Wiley-VCH GmbH