FERMat: Foundational Representation of Materials

The FERMat project aims to establish a new technological paradigm and the software infrastructure necessary for the development of Machine Learning (ML) models capable of predicting the properties of unseen molecular and materials systems/structures, thus accelerating the computational discovery of new molecules and materials. We will discuss progress in this direction, including the expansion and revised standard of the ColabFit Exchange, the largest public repository of first-principles data standardized for training ML interatomic potentials (MLIPs), and its integration with Intel's MatSciML Toolkit. Additionally, we will discuss KLIFF, a platform for training MLIPs at scale from multiple datasets in ColabFit and deploying MLIPs to popular simulation packages such as LAMMPS and ASE via our TorchScript-based "TorchML driver", with full support for distributed memory parallelism and GPUs. Finally, we will introduce KUSP, a package for deploying arbitrary models (e.g., implemented in JAX, TensorFlow, PyTorch) within simulation packages with minimal overhead.