Matlantis PFP v8: A Universal 96-Element Machine Learning Interatomic Potential Featuring r2SCAN-level Accuracy

Universal machine learning interatomic potentials (MLIPs) are transforming materials science by providing the quantum-level accuracy essential for novel scientific discovery and the computational efficiency required for large-scale industrial applications. Here, we introduce Matlantis’s latest results in developing a universal MLIP, PreFerred Potential (PFP) version 8. PFP v8 has been trained on large-scale DFT datasets, including a PBE-GGA dataset covering arbitrary combinations of 96 elements, and an r2SCAN dataset covering 70 elements. Our DFT datasets include over 60 million DFT-calculated structures to ensure informativeness, representativeness, and diversity across bulk materials (both structurally and compositionally) and molecules under a wide range of thermodynamic conditions. The extensive training of PFP v8 took more than 2,000 GPU years which underpins its improved energy/force accuracy: e.g., the testing MAEs on energy and force are 0.0023 eV/atom and 0.0017 eV/Å, respectively. Furthermore, our preliminary benchmarks on the application of PFP v8 in real-world scenarios show a significantly improved predictive capability when compared with experimental results. For example, PFP v8 with the r2SCAN mode demonstrates more accurate predictions for phase transitions (even for the phase transition temperatures) of ferroelectric materials (e.g., BaTiO3),complex structural/chemical evolutions for cathode materials in typical cycling process, and the densities of typical (organic) liquids at room temperature, just to name a few. Finally, it is possible to run LAMMPS calculations with PFP as the force field in the Matlantis environment. This work underscores PFP’s potential to significantly enhance the accuracy and scope of materials simulations.