Implicit Machine Learning Force Fields Speed Up Molecular Dynamics Simulations

A novel method known as implicit machine learning force fields (I-MLFFs) has been developed by researchers, substituting explicit neural network layers with self-consistent fixed-point equations. This innovative formulation permits the reuse of intermediate representations throughout successive timesteps in molecular simulations, facilitating warm-starting for force evaluations. The models produced merge the efficiency of a shallow, single-layer MLFF with the accuracy and representational power of a deep neural network. This technique realizes efficiency improvements that are architecture-agnostic, which are unattainable when force prediction and trajectory integration are treated independently. Tested on three primary categories of graph neural networks—SO(3)-equivariant spherical-tensor, equivariant Cartesian tensor, and invariant architectures—the method achieves a two- to five-fold decrease in computational and memory requirements while maintaining accuracy. The research can be found on arXiv with the identifier 2607.29158.