With the development of universal machine learning interatomic potentials, a rapidly growing number of chemical space data sets have appeared. One of the biggest challenges is that these data sets are mostly generated at different quantum chemical (QC) levels. However, a general framework that is scalable to learning across both the chemical space and quantum chemical levels remains unmet. In this work, we propose an all-in-one approach that 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. We showcase the superiority of our all-in-one strategy by creating OMNI-P1─the first-ever universal interatomic potential capable of simultaneously learning and making predictions at different QC levels. The generalization capability of the universal model OMNI-P1 for organic molecules is comparable to semiempirical GFN2-xTB and common density functional theory (DFT) methods with a double-ζ basis set, while its speed is orders of magnitude faster. Due to its unique ability to make predictions at different levels, a single model trained with our approach provides a straightforward way to also generate correction terms. This can be used in Δ-learning models without the need to train a dedicated correction model. We utilized this capability of OMNI-P1 to correct the DFT ωB97X-D4 level to obtain the Ω-ωB97X-D4 method with superior accuracy.
Chen et al. (Fri,) studied this question.