Reproduce “Implicit Machine Learning Force Fields Accelerate Molecular Dynamics Simulations” on OpenResearch

About this paper

Johannes Maeß, Leon Werner, J. Thorben Frank, Winfried Ripken, Martin Michajlow, Joshua Futterer, Klaus-Robert Müller, Stefan Chmiela

The authors introduce an implicit force field that reuses representations across MD steps. One neural layer is iterated to a fixed point, while extrapolation warm-starts representation and force solves. On MD17, implicit PaiNN reaches force MAE 0.37 versus 0.75 kcal/mol/Å for explicit PaiNN at matched layer-call cost; higher-compute explicit models may be more accurate, and difficult states may need more iterations.

Key insights

• Recasting force prediction into a self-consistent fixed-point problem allows MLFFs to exploit the temporal continuity of MD trajectories, amortizing computational costs across successive timesteps.
• Implicit models can achieve comparable or superior force accuracy to deeper explicit networks with a computational cost and memory footprint equivalent to a shallow, single-layer model.
• Warm-starting the fixed-point solvers from previous timesteps drastically reduces the average number of iterations required per MD step, enabling most conformations to converge in a single iteration.

Summary by alphaXiv.