A machine learning-oriented pseudo-field approach to accelerate runtime of molecular dynamics simulation of liquids

Machine learning methods are increasingly used in research to save time and computational expenses. The data-driven approach relies on existing data for predictions based on statistical inference and interpolation. Liquids, in particular, benefit from machine learning due to their higher computational requirements. In this study, we propose a simple data-driven strategy that employs classical molecular dynamics simulations to generate potential energy spatial distribution data. This data is then used to train a machine learning model by using topologically similar scaled-down systems. The trained model is subsequently employed to predict force field behaviour for complex full-scale systems, resulting in time and computational cost savings. During training, the potential energy pseudo field serves as a model descriptor, while the final predictions are generated using multidimensional regression-based learning (Gaussian Process) and neural architecture learning (Artificial Neural Network), which are integrated into the simulation as lookup tables. Comparing the data-driven and conventional approaches reveals a significant acceleration in overall simulation runtime when appropriately trained machine learning models are utilised. Although the initial training phase of the machine learning models is time-consuming, retraining is unnecessary for future simulations with the same setup. Thus, this approach offers a straightforward means of conducting complex simulations in less time.