Accelerated Development of ReaxFF Forcefields Using Machine Learning

ReaxFF based reactive force field approach has increased the application of reactive molecular dynamics simulation for studying complex material properties and processes. However, complex simulations require careful tuning of model parameters to high quality ground truth data obtained from accurate quantum mechanical methods. Also, a parameter set tuned for a particular system can be extended through optimization to other applications for similar systems since ReaxFF uses fixed functional form representation of all chemical systems. This optimization process is challenging due to large parameter space, especially for multi component chemical systems and therefore, is extremely challenging to achieve. Currently popular local (parabolic interpolation methods, gradient based methods) and global black box optimization methods (Genetic Algorithms, Monte Carlo Methods, Covariance Matrix Adaptation evolution strategy, Particle Swarm Optimization) require hundreds of thousands of error evaluations for complex training tasks and subsequent error evaluation through energy minimization of many several molecules in the training set. Therefore, an efficient optimization scheme is needed for sampling this high dimensional parameter space with many local minima. Here, we demonstrate various machine learning based parameter optimization scheme, that use a parallel MD agent for environment sampling and utilizing various machine learning methods to efficiently optimize high dimensional parameter sets. These parameter sets can be further fine-tuned by using gradient based gradient approached near local minima to yield high quality parameter sets suitable for a particular problem at hand<br/><br/><b><u>Acknowledgments</u></b><br/>This work was supported as part of the Computational Materials Sciences Program funded by the<br/>U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award Number<br/>DE-SC001460. Simulations were performed at the Argonne Leadership Computing Facility<br/>under the DOE INCITE and Aurora Early Science programs and at the Center for Advanced<br/>Research Computing of the University of Southern California