Pretraining Strategy for Neural Potentials

Molecular dynamics plays a pivotal role in elucidating the dynamic behavior of molecules, providing essential insights into the temporal evolution of complex systems, such as understanding the conformational changes, interactions, and thermodynamic properties of molecules. It is thus important for fields ranging from drug discovery to materials science. The potential energy surface serves as the underlying landscape dictating the dynamics of a molecular system, and making an accurate representation of it is crucial for studying the dynamics of molecules through numerical simulations. Ab initio methods like Density Functional Theory (DFT) provide high accuracy by accounting for the electronic structure of atoms. Yet their high computational demands pose a significant challenge for efficient utilization, particularly in the context of large many-body systems. On the other hand, forces can be computed using empirical interatomic potentials tailored to specific environments, bypassing the electronic structures of the system. This approach significantly reduces the computational cost compared to ab initio methods. As there is a wide variety of interactions in molecular systems (bonded or non-bonded interactions), finding the appropriate functional forms can be challenging.