Title:Beyond Independent Error Assumptions in Large GNN Atomistic Models

Abstract:The practical applications of determining the relative difference in adsorption energies are extensive, such as identifying optimal catalysts, calculating reaction energies, and determining the lowest adsorption energy on a catalytic surface. Although Density Functional Theory (DFT) can effectively calculate relative values through systematic error cancellation, the accuracy of Graph Neural Networks (GNNs) in this regard remains uncertain. To investigate this issue, we analyzed approximately 483 million pairs of energy differences predicted by DFT and GNNs using the Open Catalyst 2020 - Dense dataset. Our analysis revealed that GNNs exhibit a correlated error that can be reduced through subtraction, thereby challenging the naive independent error assumption in GNN predictions and leading to more precise energy difference predictions. To assess the magnitude of error cancellation in chemically similar pairs, we introduced a new metric, the subgroup error cancellation ratio (SECR). Our findings suggest that state-of-the-art GNN models can achieve error reduction up to 77% in these subgroups, comparable to the level of error cancellation observed with DFT. This significant error cancellation allows GNNs to achieve higher accuracy than individual adsorption energy predictions, which can otherwise suffer from amplified error due to random error propagation.