Implicit ML Force Fields Boost Molecular Simulation Speed, Accuracy

This paper presents Implicit Machine Learning Force Fields (I-MLFFs), a novel method designed to dramatically accelerate molecular dynamics simulations. Unlike traditional MLFFs that use explicit neural network layers, I-MLFFs replace these with self-consistent fixed-point equations. This architectural change allows for the reuse of intermediate representations across successive timesteps, effectively "warm-starting" force evaluations. The key benefit of I-MLFFs is their ability to combine the computational efficiency of shallow, single-layer MLFFs with the high representational capacity and accuracy typically found in deep neural networks. This leads to architecture-agnostic efficiency gains, meaning it works across various graph neural network types, including invariant, equivariant Cartesian tensor, and SO(3)-equivariant spherical-tensor architectures. The results show a two- to five-fold reduction in compute and memory footprint. Crucially, these improvements are achieved without sacrificing atomistic resolution or the original integration timestep, avoiding the need for spatial or temporal coarse-graining. This breakthrough significantly extends the capabilities of quantum-mechanically faithful molecular simulations, enabling larger systems and longer trajectories within existing hardware constraints, which could unlock new insights in biomolecular and material sciences.