(152e) Machine-Learning a Solution for Reactive Simulations of Complex Chemical Systems | AIChE

(152e) Machine-Learning a Solution for Reactive Simulations of Complex Chemical Systems


Lindsey, R. - Presenter, Lawrence Livermore Nat'L Lab.
Fried, L. E., Lawrence Livermore National Laboratory
Goldman, N., Lawrence Livermore National Laboratory
Pham, C. H., Lawrence Livermore Nat'l Lab
Bastea, S., Lawrence Livermore National Laboratory
Many grand challenges in computational chemistry involve accurate description of condensed-phase reactivity. For example, understanding chemistry at extreme conditions (e.g. 1000s of K and 10s of GPa) is crucial in fields spanning geochemistry, astrobiology, and alternative energy. Quantum mechanical methods like Kohn-Sham density functional theory (DFT) can provide valuable microscopic insights into such systems while circumventing the risks of physical experiments, but the associated reactive scales (i.e. ns and μm) largely preclude extension of such models to molecular dynamics. These limitations have motivated extensive work towards development of scalable classical reactive interatomic potentials; however, the complicated forms underlying these models require time consuming parameterization, yet generally still fail to capture the physics of chemistry under extreme conditions. As a result, the physiochemical space over which accurate reactive interatomic potential parameters exist is limited.

In this work, we present the Chebyshev Interaction Model for Efficient Simulation (ChIMES), a machine-learned generalized many-body reactive interatomic potential. ChIMES models are constructed from linear combinations of Chebyshev polynomials, and are rapidly generated by force, stress, and energy matching to short DFT trajectories. These models retain most of the accuracy of DFT while decreasing computational requirements by several orders of magnitude and enabling linear-scaling with respect to system size. To date, these models have been applied to condensed-phase reaction-driven problems including shockwave-driven nano-carbon formation, hydride embrittlement, and equation-of-state prediction. Details of the model and fitting approach will be discussed along with selected applications.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS- 808405