Grateful to everyone involved: @thorbenfrank.bsky.social, Sergio S. Dou, Almaz Khabibrakhmanov, Leonardo M. Sandonas, Oliver T. Unke, Stefan Chmiela, Klaus-Robert Müller, and Alexandre Tkatchenko.

This has been an enjoyable collaborative effort between @uni.lu, @tuberlin.bsky.social/@bifold.berlin, and Google DeepMind.

We hope this contribution will enable new simulations&insights and be of value to the community working on the next generation of general-purpose MLFFs.

The model, code and data are available here: github.com/general-mole...

To assess its capabilities and limitations, we benchmarked SO3LR on systems ranging from small molecules to large solvated biomolecules: protein, glycoprotein, and lipid bilayer. The model proved stable, reproduced local and global structural properties, and scaled to ~200k atoms on a single GPU.

It unites an equivariant neural network for semi-local effects with explicit physical potentials for short-range repulsion and long-range electrostatics/dispersion. The model was trained in <100 GPUh on a curated dataset of 4M molecular systems computed at the PBE0+MBD level of theory.

A persistent challenge in atomistic modelling is combining quantum-level accuracy with the efficiency needed for large, complex simulations. Our pretrained lightweight SO3LR MLFF is a step toward this goal.