Molly.jl: Machine Learning Potentials for Molecular Dynamics

Classical force fields in molecular dynamics are fast but cannot capture bond breaking, charge transfer, or complex many-body chemistry. Machine learning potentials (MLPs) close this gap — offering near-quantum accuracy at a fraction of ab initio cost. Molly.jl currently supports MLPs only through a Python bridge (ASECalculator), which requires a Python runtime and bypasses Julia's differentiable programming stack entirely. This project delivers native Julia support for four widely-used MLP families: ANI (neural network potential for organic molecules), ACE (linear body-ordered basis expansion), NequIP, and Allegro (E(3)-equivariant graph neural networks). ANI is implemented from scratch in pure Julia using Lux.jl, with pre-trained ANI-1x/ANI-2x weights loaded from HDF5. ACE is integrated via the mature ACEpotentials.jl package. NequIP and Allegro are wrapped from TorchScript-exported models via PythonCall, following the existing ASECalculator pattern. All four potentials plug into Molly.jl's general_inters dispatch loop with zero changes to core simulation code. Pure-Julia models (ANI, ACE) are fully differentiable via Enzyme reverse-mode AD and GPU-accelerated via KernelAbstractions. The result: researchers can run differentiable, ML-accelerated molecular dynamics entirely within Julia — no Python, no overhead, composable with the full SciML ecosystem.