Multi-fidelity Approach to Data Efficient Construction of Graph Neural Network Interatomic Potentials

The development of machine learning interatomic potentials (MLIPs) for atomistic simulations has attracted increasing attention, due to their near ab initio accuracy at a fraction of computational cost. The actual quality of these interatomic potentials is constrained by the accuracy of reference methods. Density functional theory (DFT) at the level of generalized-gradient approximation (GGA) functionals is commonly used for reference data generation because of its favorable trade-off between accuracy and efficiency. However, recently developed meta-GGA and hybrid functionals have demonstrated a significant improvement over GGA functionals in describing atomic interactions. This results in more reliable predictions for the structural and dynamical properties of complex systems, albeit with an associated increase in computational cost. In this work, we present an efficient multi-fidelity approach to construct highly accurate graph deep learning interatomic potentials using a minimal amount of expensive beyond-GGA calculations. We demonstrate the power of this approach on a series of benchmark systems with different types of chemical bonding.