American Chemical Society
jp9b06348_si_002.mp4 (1.26 MB)

Teaching an Old Dog New Tricks: Machine Learning an Improved TIP3P Potential Model for Liquid–Vapor Phase Phenomena

Download (1.26 MB)
posted on 2019-09-03, 13:35 authored by Troy D. Loeffler, Henry Chan, Kiran Sasikumar, Badri Narayanan, Mathew J. Cherukara, Stephen Gray, Subramanian K. R. S. Sankaranarayanan
Water is ubiquitous yet displays a rich variety of thermodynamic properties and anomalies. An understanding of liquid–vapor phenomena in water is of broad importance to everyday processes such as evaporation, condensation, and cavitation, as well as energy technologies such as steam turbines. An accurate description of the vapor–liquid phenomena is quite challenging owing to the significant differences between how water behaves in small, sparsely distributed clusters and how it behaves in a dense bulk liquid. It is not surprising that there exist a myriad of different water models, which have attempted to describe water behavior with varying degrees of success. In general, water models have evolved from simple three-point transferable interatomic potentials (TIP3P) to more complex four-point and five-point TIP models to more recent polarizable models. The natural evolution from TIP3P to TIP4P families of models was, in part, due to the belief that we have perhaps reached the limit of what the simple three-point models are capable of achieving. The advent of big data analytics and ever-increasing supercomputing resources has brought to the forefront powerful machine learning techniques for materials design. Here, we take advantage of machine learning techniques such as hierarchical objective genetic algorithms to demonstrate that simple computationally efficient models developed decades ago can be retrained to perform significantly better than their original counterparts. In a departure from typical practice, we train our model against an elaborate temperature-dependent data obtained from molecular dynamics trajectories to cluster properties using extensive configurational sampling and on-the-fly Monte Carlo simulations. To demonstrate the power of our machine learning approach, we choose the popular TIP3P model that, however, is widely acknowledged to perform poorly in describing vapor–liquid properties. We retrain this TIP3P model to dramatically improve its performance over the original model. Our new ML-TIP3P performs on par or, in some respects, better than even the current best performing nonpolarizable model (TIP4P/2005) for vapor–liquid properties. To exemplify the suitability of our approach, we apply our newly developed model to study a highly nonequilibrium vapor–liquid phenomenon, laser-induced heterogeneous cavitation in a gold–water system. Overall, our study highlights a general strategy for atomistic model development that can be potentially used to retrain existing potential models and help them attain their best possible performance.