A Transferrable Range-Separated Force Field for Water: Combining the Power of Both Physically-Motivated Models and Machine Learning Techniques

An accurate, transferrable, and computationally efficient potential energy surface (PES) is of paramount importance for all molecular mechanics simulations. In this work, using water as example, we demonstrate how one can construct a reliable force field by combining the advantages of both physically-motivated and data-driven machine learning (ML) methods. Different to the existing water models based on many-body expansion, we adopt a separation scheme purely based on distances, and systematically investigate how the long-range asymptotic terms increase the transferability and the data efficiency of the ML potential. We also show how the ML model can be an ideal tool to fit the short-range interactions which used to post great challenges to the conventional physically-motivated force fields. The water force field we obtain is highly accurate and transferrable in different environments, and the distance-based separation scheme is easy to be extended to general molecular systems. Through this study, we show how the information we learn from small clusters can be extrapolated into larger systems, thus providing a general recipe for the intermolecular force field development at CCSD(T) level of theory in future.