Advancing predictive molecular modeling: Accuracy and stability of physics-based molecular many-body models

The MB-pol water model, designed to correctly capture the physics of molecular many-body interactions, has been shown to correctly predict the phase diagram of water. We present our studies advancing the MB-nrg theoretical/computational framework both in terms of accuracy and on stability.

We present the research findings of the paper "MB-pol(2023): Sub-Chemical Accuracy for Water Simulations from the Gas to the Liquid Phase." Our study focuses on the development of a new family of data-driven many-body potential energy functions (PEFs) for water, named MB-pol(2023). Importantly, the MB-pol(2023) PEFs exhibit remarkable agreement with experimental results for various properties of liquid water, effectively closing the gap with experimental measurements. The improved accuracy of the MB-pol(2023) PEFs in characterizing the structural, dynamic, and reactive properties of liquid water highlights their significant potential for a wide range of applications.

We also present our progress in improving the stability of the MB-nrg molecular modeling framework. Under extreme thermodynamic conditions, machine-learning-based molecular models sometimes exhibit numerical instability due to unpredictable behaviors when encountering input configurations significantly different from the training configurations. We demonstrate that the physical nature of the molecular many-body expansion can be leveraged to select mathematical formulations that guarantee stability.