Multiobjective genetic training and uncertainty quantification of reactive force fields

npj Computational Materials, Aug 2018

The ReaxFF reactive force-field approach has significantly extended the applicability of reactive molecular dynamics simulations to a wide range of material properties and processes. ReaxFF parameters are commonly trained to fit a predefined set of quantum-mechanical data, but it remains uncertain how accurately the quantities of interest are described when applied to complex chemical reactions. Here, we present a dynamic approach based on multiobjective genetic algorithm for the training of ReaxFF parameters and uncertainty quantification of simulated quantities of interest. ReaxFF parameters are trained by directly fitting reactive molecular dynamics trajectories against quantum molecular dynamics trajectories on the fly, where the Pareto optimal front for the multiple quantities of interest provides an ensemble of ReaxFF models for uncertainty quantification. Our in situ multiobjective genetic algorithm workflow achieves scalability by eliminating the file I/O bottleneck using interprocess communications. The in situ multiobjective genetic algorithm workflow has been applied to high-temperature sulfidation of MoO3 by H2S precursor, which is an essential reaction step for chemical vapor deposition synthesis of MoS2 layers. Our work suggests a new reactive molecular dynamics simulation approach for far-from-equilibrium chemical processes, which quantitatively reproduces quantum molecular dynamics simulations while providing error bars.

A PDF file should load here. If you do not see its contents the file may be temporarily unavailable at the journal website or you do not have a PDF plug-in installed and enabled in your browser.

Alternatively, you can download the file locally and open with any standalone PDF reader:

https://www.nature.com/articles/s41524-018-0098-3.pdf

Multiobjective genetic training and uncertainty quantification of reactive force fields

Abstract The ReaxFF reactive force-field approach has significantly extended the applicability of reactive molecular dynamics simulations to a wide range of material properties and processes. ReaxFF parameters are commonly trained to fit a predefined set of quantum-mechanical data, but it remains uncertain how accurately the quantities of interest are described when applied to complex chemical reactions. Here, we present a dynamic approach based on multiobjective genetic algorithm for the training of ReaxFF parameters and uncertainty quantification of simulated quantities of interest. ReaxFF parameters are trained by directly fitting reactive molecular dynamics trajectories against quantum molecular dynamics trajectories on the fly, where the Pareto optimal front for the multiple quantities of interest provides an ensemble of ReaxFF models for uncertainty quantification. Our in situ multiobjective genetic algorithm workflow achieves scalability by eliminating the file I/O bottleneck using interprocess communications. The in situ multiobjective genetic algorithm workflow has been applied to high-temperature sulfidation of MoO3 by H2S precursor, which is an essential reaction step for chemical vapor deposition synthesis of MoS2 layers. Our work suggests a new reactive molecular dynamics simulation approach for far-from-equilibrium chemical processes, which quantitatively reproduces quantum molecular dynamics simulations while providing error bars. Introduction The reactive molecular dynamics (RMD) method has enabled large-scale simulations of chemical events in complex materials involving multimillion atoms.1,2 In particular, RMD simulations based on first principles-informed reactive force fields (ReaxFF)3 describe chemical reactions (i.e., bond breakage and formation) through a bond-order/distance relationship that reflects each atom’s coordination change. ReaxFF–RMD simulations describe full dynamics of chemical events at the atomic level with significantly reduced computational cost compared with quantum-mechanics (QM) calculations.4 ReaxFF consists of a number of empirical force-field parameters in its functional form, which are optimized mainly against a QM-based training set using a single-parameter parabolic search scheme.5 In addition to such a well-established optimization technique, several ReaxFF optimization frameworks have been developed recently using multiobjective genetic algorithms (MOGA) and other evolutionary optimization methods.6,7 QM data points in a training set include not only energies of small clusters (e.g., full bond dissociation, angle distortion and torsion energies) and reaction energies/barriers for key chemical reactions, but also bulk properties of crystal systems (e.g., equations of state, bulk modulus and cohesive energies).8,9 As a result, ReaxFF has shown its ability to successively study chemical, physical and mechanical properties of a wide range of complex materials such as hydrocarbon,10 high energy materials11 and metal/transition-metal systems.12,13 Despite these successes, the transferability of ReaxFF to highly non-equilibrium processes such as high-temperature reactions remains to be established. This is because the QM data points used in force-field optimization are mainly static quantities like ground/intermediate/transition-state structures and energies. For more accurate RMD simulations of far-from-equilibrium reaction dynamics, we here propose a dynamic approach, where ReaxFF parameters are calibrated by directly fitting RMD trajectories against quantum molecular dynamics (QMD)14,15,16,17 trajectories on the fly. This dynamic approach is implemented using MOGA to optimize the ReaxFF model in terms of multiple quantities of interest (QoI). MOGA uses a non-dominated sorting algorithm18,19 to sort out a population of ReaxFF models into different sets called Pareto optimal fronts, such that every set contains models non-dominated by each other in terms of accuracy of the QoI. Further information regarding the implementation of this algorithm is provided in the Methods section and Supplementary Information. While significantly extending the applicability of ReaxFF to a wider range of reactions and processes, the above dynamic approach poses a challenge in estimating uncertainties in the force-field parameters and propagating them to those in model predictions. Such uncertainty quantification (UQ) has become central to most computational sciences.20 In particular, Bayesian-ensemble approaches have been applied successfully to UQ of force-field parameters21,22 in molecular dynamics (MD) simulations and the exchange-correlation functional in density functional theory (DFT) for QMD simulations.23,24 These approaches are typically employed within the conventional training of model parameters so as to minimize a weighted sum of errors against given ground-truth values in a training set. While multiobjective training like MOGA makes the application of standard (...truncated)


This is a preview of a remote PDF: https://www.nature.com/articles/s41524-018-0098-3.pdf

Ankit Mishra, Sungwook Hong, Pankaj Rajak, Chunyang Sheng, Ken-ichi Nomura, Rajiv K. Kalia, Aiichiro Nakano, Priya Vashishta. Multiobjective genetic training and uncertainty quantification of reactive force fields, npj Computational Materials, 2018, Issue: 4, DOI: 10.1038/s41524-018-0098-3