OECD bibliometric indicators: Selected highlights, April 2024
Development of a Machine Learning Algorithm for High Energy Density Materials
1. Development of a Machine Learning
Algorithm for High Energy Density
Materials
Svetlana Gelpi
Major advisor: Dr. José A. Gascón
General Examination
September 17, 2019
5. Introduction: Applications
Commercial
IndustrialMilitary
Goll, James G., L. Jenifer, and M. Tarah. "Teaching chemistry using the girls with yellow hands." Journal
Chemical educator 13 (2008): 3-5.
https://www.thecarconnection.com/news/1079790_phony-airbags-what-to-do-if-you-have-them-
on-your-car
https://cen.acs.org/articles/95/i27/s-fireworks-produces-those-colorful.html
https://www.ausimmbulletin.com/feature/developing-alternative-explosives-blasting-without-nitrogen-
oxide-hazard/
5
• Insensitivity.
• Moldability (mixtures).
• Increase in detonation performance.
• Environmental impact.
Variables to improve on:
7. Motivation
Energetic material discovery Characterization Detonation trials Agency approval
5-10 years
• Expensive facilities.
• Only small amounts can be
used.
• Trial and error.
7
8. Motivation
Energetic material discovery Characterization Detonation trials Agency approval
5-10 years
• Structure.
• Density.
• Sensitivity.
• Heat of reaction.
Men, Z.; Suslick, K. S.; Dlott, D. D. The Journal of Physical Chemistry C 2018, 122, 14289. 8
9. Motivation
Energetic material discovery Characterization Detonation trials Agency approval
5-10 years
• Dynamic testing.
• Specialized
equipment.
• Limited runs per day.
• Theoretical codes.
9
10. Motivation
Energetic material discovery Characterization Detonation trials Agency approval
10-12 years
In the United States of America:
• Department of Defense
(The Pentagon).
• Bureau of alcohol, tobacco,
firearms, and explosives.
• Department of
Transportation.
10
5-10 years
13. Problem statement
There exists a need to predict fast and reliable
heats of formation for the development of new
explosives.
13
14. Previous research
QM Atomization energies + Isodesmic reactions
∆𝐻#
°
𝐴& 𝐵(, 0𝐾 = 𝑥∆𝐻#
°
A, 0K + y∆𝐻#
°
𝐵, 0𝐾 − 3 𝐷5
Ochterski, J. W. Thermochemistry in Gaussian; Gaussian Inc.: Wallingford, CT, 2000.
Machine Learning
Elton, D. C.; Boukouvalas, Z.; Butrico, M. S.; Fuge, M. D.; Chung, P.
W. Scientific Reports 2018, 8, 9059.
Ubaru, S.; Międlar, A.; Saad, Y.; Chelikowsky, J. R. Physical Review
B 2017, 95, 214102.
14
1.
2. 3.
3
Group Additivity Methods
+ 2 + 1 =
Becerra, R.; Walsh, R. Physical Chemistry Chemical
Physics 2019, 21, 988.
15. Previous attempts at predicting
heats of formation using ML
Machine Learning
Elton, D. C.; Boukouvalas, Z.; Butrico, M. S.; Fuge, M. D.; Chung, P. W.
Scientific Reports 2018, 8, 9059.
Ubaru, S.; Międlar, A.; Saad, Y.; Chelikowsky, J. R. Physical
Review B 2017, 95, 214102.
Method: Kernel Ridge Regression.
Featurizations: Descriptors and Coulomb
Matrices.
Results: R2 = 0.94.
Limitations: Transferability.
Method: Support Vector Regression.
Featurizations: Experimental data.
Results: R2 = 0.87.
Limitations: Trained on metal alloys,
transferability.
15
16. ANAKIN-ME (ANI-1)
• Trained on drug-like molecules.
• Models DFT with high precision.
• Neural network potential based ML.
Smith, J. S.; Isayev, O.; Roitberg, A. E. Chemical Science 2017, 8, 3192. 16
21. Proposal
1.Re-train the ANI ML algorithm to predict total
quantum energies for explosive molecules.
2.Train a ML algorithm to predict heats of
formation.
21
Specific Aims:
22. Is re-training necessary?
Performs well on drug like molecules but what happens when
you are in a different sub chemical space?
22
Specific Aim 1: Re-train the ANI ML algorithm to predict total quantum energies
for explosive molecules.
Smith, J. S.; Isayev, O.; Roitberg, A. E. Chemical Science 2017, 8, 3192.
23. Correlation total quantum energies (TQE)
Goal: Compare TQE obtained from ANI and DFT
23
10 crystal structures from
Huang and Massa dataset.
Optimize structures using
wb97xd/6-31g* (DFT).
Predict TQE using ANI. Calculate TQE using DFT .
25. Preliminary results
Conclusion: The data produced for the ten molecules does not serve
as a valuable metric to evaluate the predictive capabilities of ANAKIN.
25
26. Preliminary results: a closer look
NOEURA
Conformational sampling performed:
• 3.9 ps simulation performed at 300K with NVE ensemble
Image from Heather Kulik (MIT) via OpenCourseWare.
26
More stringent test
Oxygen
Nitrogen
Hydrogen
Carbon
27. Preliminary results
27
Extract 100 conformations (E1, E2,
E3, ….. E100). Calculate E for DFT
and ANI algorithm.
From sampling, extract
lowest energy
conformation (E0).
Find ∆E for wb97xd and ANI.
Compare ∆Ewb97xd
and ∆EANI.
28. Preliminary results
Linear correlation of single point energy calculations on
conformations.
28
-5 0 5 10 15 20 25 30
-5
0
5
10
15
20
25
30
-5 0 5 10 15 20 25 30
-5
0
5
10
15
20
25
30
PEP[kcal/mol]
ANI-1
PEP (Benchmark (wb97xd) [kcal/mol])
R2=0.65 standard error of 3.54 kcal/mol.
30. Specific Aims
1.Re-train the ANI ML algorithm to predict total
quantum energies for explosive molecules.
ü Molecular Dynamics simulations of
explosive mixtures.
2.Train a ML algorithm to reproduce heats of
formation.
ü Use ΔHf° in thermochemical codes to
predict explosives performance.
30
31. Difference in Complexity TQE and ΔHf°
31
Ground State TQE ΔHf°
𝐻Ψ = 𝐸Ψ
Microstate model
0K Temperature
Macrostate model
298K Temperature
Temperature contributes to the total energy.
32. Specific Aim 2 – Training a neural
network
𝑀𝑋7
𝑀𝑋<
𝑀𝑌>
AEV
NN
Prediction
𝐸 𝑋
1
𝐸 𝑋
2
𝐸 𝑌
3
ΔHf°
Atomic
Contribution
Training with Huang and Massa explosives dataset
32Huang, L.; Massa, L. “APPLICATIONS OF ENERGETIC MATERIALS BY A THEORETICAL METHOD (DISCOVER ENERGETIC MATERIALS
BY A THEORETICAL METHOD),” 2013.
33. Conclusions
• Re-training of the ANI algorithm will help facilitate
long molecular dynamics simulations which will help
evaluate the properties of explosives in mixtures.
• Training of a NN algorithm will help facilitate the
prediction of fast and reliable heats of formation for
explosive molecules.
33
34. Acknowledgements
34
Amarilys Domínguez
Dr. Angelo Rossi
Dr. Sergio Wong
Dr. Brian J. Bennion
Dan Kirshner
Dr. Terianna Wax
Dr. Alfredo Angeles-Boza
Alyssa Hartmann
Sonia Chavez
Juanita Ordoñez
Rafael Rivera
Shubhashish Verma
Dr. Natalia Pigni
Alissa Richard
Mansi Malhotra
Kevin Clark
Laura Achola
Karla Arias
Luisa Posada
Dr. Ehsan Moharreri
Sameera Sansare
Koyel Sen
Committee Members:
Dr. José A. Gascón
Dr. Jessica Rouge
Dr. Fatma Selampinar
Dr. Christian Brückner
Dr. Steven L. Suib
41. Bond energies
41
C-C single bond = 83 kcal/mol
C=C double bond = 146 kcal/mol
C#C triple bond = 200 kcal/mol
C-N single bond = 73 kcal/mol
C=N double bond = 147 kcal/mol
C#N triple bond = 213 kcal/mol
N-N single bond = 38 kcal/mol
N=N double bond = 109 kcal/mol
N#N triple bond = 226 kcal/mol
O-N-O single bond = 72 kcal/mol
O2N-O single bond = 45 kcal/mol
Lower bond energy = less bond stability
42. How are bond energies
experimentally determined?
42
What energy is necessary to break a bond?
𝐸 =
ℎ𝑐
𝜆
IR spectroscopy: vibration of a given bond is proportional to the bond dissociation energy.
Atomizing a sample and studying the emission.
45. Relationship ∆Hf and detonation
properties
45
A higher performance for secondary explosives is of
fundamental importance and
always desired. The main performance criteria are:
1. heat of explosion Q
2. 2. detonation velocity D
3. 3. detonation pressure P,
and less importantly,
4. explosion temperature T (K)
5. volume of gas released V per kg explosive .
47. Perform MD simulations
Once the algorithm has been trained and tested, MD simulations of a recently developed
explosive ICM-10112 similar to that of other explosives that need to be stabilized in mixtures
will be performed. Simulations of the isolated and polymer mixtures will be carried out at
the NVT ensemble at 298K using isolated ICM-101, ICM-101 with PEG, ICM-101 with
EVA, and ICM-101 with F2311 for an initial 100 ps with Periodic Boundary Conditions.11
Mechanical properties, binding energies, and moldability will be analyzed from the MD
simulation trajectories. These results will give insight into how the explosive molecules
properties are enhanced or decreased. This will help in the
Yuehai Yu, RSC Adv., 2016,6, 20034-20041
47
48. Reaction Product Hierarchy
48
Rules of thumb:
1) All nitrogen becomes N2.
2) All available oxygen goes first to convert hydrogen to H2O.
3) Any oxygen left after H2O formation burns carbon to CO.
4) Leftover oxygen from step (3) converts CO to CO2.
5) Leftover oxygen from step (4) forms O2 and is available for
use in secondary reactions.
6) Traces of Nox are always formed.
7) Any leftover carbon becomes solid residue (graphite).
Pages 11-15 B. Lusk and J. J. Silva Energy distrubtion in the blast fragmentation process.
50. Heats of formation using Isodesmic
Reactions
50
• No change in number of each bond type between
reactants and products
Ex: CO2 + CH4 ® 2 H2CO
(2 C=O and 4 C-H bonds on each side)