Data-efficient multi-fidelity training for high-fidelity machine\n learning interatomic potentials

Machine learning interatomic potentials (MLIPs) are used to estimate\npotential energy surfaces (PES) from ab initio calculations, providing near\nquantum-level accuracy with reduced computational costs. However, the high cost\nof assembling high-fidelity databases hampers the application of MLIPs to\nsystems that require high chemical accuracy. Utilizing an equivariant graph\nneural network, we present an MLIP framework that trains on multi-fidelity\ndatabases simultaneously. This approach enables the accurate learning of\nhigh-fidelity PES with minimal high-fidelity data. We test this framework on\nthe Li$_6$PS$_5$Cl and In$_x$Ga$_{1-x}$N systems. The computational results\nindicate that geometric and compositional spaces not covered by the\nhigh-fidelity meta-gradient generalized approximation (meta-GGA) database can\nbe effectively inferred from low-fidelity GGA data, thus enhancing accuracy and\nmolecular dynamics stability. We also develop a general-purpose MLIP that\nutilizes both GGA and meta-GGA data from the Materials Project, significantly\nenhancing MLIP performance for high-accuracy tasks such as predicting energies\nabove hull for crystals in general. Furthermore, we demonstrate that the\npresent multi-fidelity learning is more effective than transfer learning or\n$\\Delta$-learning an d that it can also be applied to learn higher-fidelity up\nto the coupled-cluster level. We believe this methodology holds promise for\ncreating highly accurate bespoke or universal MLIPs by effectively expanding\nthe high-fidelity dataset.\n