One To Rule Them All: A Universal Interatomic Potential Learning Across Quantum Chemical Levels

All so far reported universal ML interatomic potentials target a single quantum chemical (QC) level due to the challenges associated with learning on heterogeneous data sets. Here we introduce an all-in-one approach for learning across QC levels, which we leverage to create OMNI-P1 – the first-ever universal interatomic potential simultaneously trained on different levels. Our all-in-one approach enables simultaneous learning on an arbitrary number of QC levels from various data sets, presenting a more general and easier-to-use alternative to transfer learning. The generalization capability of the universal model OMNI-P1 for organic molecules is comparable to semi-empirical GFN2-xTB and common density functional theory (DFT) methods with a double-zeta basis set, while the speed is orders of magnitude faster. We have also built the universal Δ-learning model OMNI-P1d that has an accuracy superior to common DFT approaches. OMNI-P1d utilizes the OMNI-P1 to conveniently obtain the corrections to lower-level predictions.