Influence of Exchange-Correlation Functional on Machine-Learned Interatomic Potentials' Accuracy: A Systematic Study...

Machine-learned interatomic potentials (MLIPs) have emerged as a strategy to accelerate molecular simulations, offering the promise of quantum chemical accuracy at a cost close to that of classical force fields. They are commonly trained on reference data obtained at the DFT level, which is itself a sufficiently affordable method to generate thousands of configurations and associated energies and atomic forces. However, this approach suffers from a major underlying challenge, which is rarely addressed: the choice of exchange-correlation functional. Although it is understood to have a crucial impact on the accuracy of the description of interactions and therefore on the results of the molecular simulations, it has not been systematically evaluated in the past. To go beyond simple benchmarks on selected configurations, we set out to understand the influence of exchange-correlation (XC) functionals on the training of MLIPs and, more importantly, on the physical properties of the resulting condensed matter systems. As a complex chemical system to test this, we selected borosilicate glasses of a wide variety of compositions. We produced reference data sets of borosilicate glasses with ten different chemical compositions, with six different XC functionals at the GGA, meta-GGA, and hybrid levels. We then trained MLIPs for all of them, produced glasses by melt-quenching, and compared their physical and structural properties with available experimental data. We show that there are significant differences between the different functionals. While the influence on bulk properties such as glass densities is minor, fine structural properties are more sensitive to the choice of functional, in particular when it comes to the coordination of boron atoms. Our results show how MLIPs can help us better understand the strengths and weaknesses of DFT functionals, where investigation through direct ab initio molecular dynamics was too expensive in the past. This also highlights that MLIPs depend on the ground truth on which they are trained, a fact often overlooked in practice, and this influence may be drastic for specific physical properties.