The MACE-MP-0a medium universal foundation model is now available in OpenKIM (DOI: 10.25950/ccbf69e7) as a Portable Model using the TorchML Model Driver. It is a Message Passing Neural Network (MPNN) trained on over 150,000 inorganic crystals, capable of zero-shot prediction for a wide range of chemical systems. The medium variant offers an optimal trade-off between computational cost and accuracy. Its publication on OpenKIM brings the plug-and-play compatibility with any KIM API compliant simulator code to this venerable interatomic potential.
Because providing the machine learning requirements for the TorchML Model Driver can be difficult on bare metal, we have also developed a version of the KIM Developer Platform Docker image with all prerequisites installed. It can be used with Docker or other container software such as Singularity/Apptainer or Podman to run TorchML Portable Models out of the box, including support for utilizing the host GPU. See the documentation here.