Data-Efficient Multifidelity\nTraining for High-Fidelity\nMachine Learning Interatomic Potentials

Machine learning interatomic potentials (MLIPs) are used\nto estimate\npotential energy surfaces (PES) from <i>ab initio</i> calculations,\nproviding near-quantum-level accuracy with reduced computational costs.\nHowever, the high cost of assembling high-fidelity databases hampers\nthe application of MLIPs to systems that require high chemical accuracy.\nUtilizing an equivariant graph neural network, we present an MLIP\nframework that trains on multifidelity databases simultaneously. This\napproach enables the accurate learning of high-fidelity PES with minimal\nhigh-fidelity data. Employing the generalized gradient approximation\n(GGA) and meta-GGA as low- and high-fidelity approaches, respectively,\nwe tested this framework on the Li<sub>6</sub>PS<sub>5</sub>Cl and\nIn<sub><i>x</i></sub>Ga<sub>1–<i>x</i></sub>N systems. The results show that using a high-fidelity training set\nwith a size approximately 10% of the low-fidelity set, the multifidelity\ntraining framework achieves excellent accuracy, with Li-ion conductivity\npredictions within 10% error and In<sub><i>x</i></sub>Ga<sub>1–<i>x</i></sub>N mixing energy showing an <i>R</i><sup>2</sup> of 0.98 compared to the reference high-fidelity\nMLIP results. It indicates that geometric and compositional spaces\nnot covered by the high-fidelity meta-GGA database can be effectively\ninferred from low-fidelity GGA data, thus enhancing accuracy and molecular\ndynamics stability. We also developed 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\nenergies above hull for crystals in general. Furthermore, we demonstrate\nthat the present multifidelity learning is more effective than transfer\nlearning or Δ-learning and that it can also be applied to learn\nhigher-fidelity up to the coupled-cluster level. We believe this methodology\nholds promise for creating highly accurate bespoke or universal MLIPs\nby effectively expanding the high-fidelity data set.