Ensemble of ∆-ML Models for Quantum Chemistry Level of Theory Selection for Interaction Energies

Ab initio quantum chemical approaches for computing interactions between molecules are useful in a wide range of applications but can be difficult to apply effectively. In particular, variations in cost and accuracy across different levels of theory (method and basis set combinations) makes it challenging to select an appropriate level of theory for a given chemical system. In this work, we train ∆-machine learning models to predict interaction energy errors for 81 levels of theory versus a reference and simple polynomial models to quantify computational cost for a meaningful subset of these methods. We demonstrate how a large language model (LLM) can leverage these models to select appropriate levels of theory and run computations autonomously on distributed resources, significantly reducing the barrier to obtaining accurate interaction energies.