Multitask methods for predicting molecular properties from heterogeneous data

• DOI: 10.1063/5.0201681
• Publication Date: 2024-07-03T12:16:57Z
• Authors: K. E. Fisher; M. F. Herbst; Y. M. Marzouk
• Keywords: Chemical Physics (physics.chem-ph); FOS: Computer and information sciences; Statistics - Machine Learning; Physics - Chemical Physics; 0103 physical sciences; FOS: Physical sciences; Machine Learning (stat.ML); 01 natural sciences

Data generation remains a bottleneck in training surrogate models to predict molecular properties. We demonstrate that multitask Gaussian process regression overcomes this limitation by leveraging both expensive and cheap data sources. In particular, we consider sets constructed from coupled-cluster (CC) density functional theory (DFT) data. report surrogates can at CC-level accuracy with reduction cost over an order of magnitude. Of note, our approach allows the set include DFT generated heterogeneous mix exchange–correlation functionals without imposing any artificial hierarchy on accuracy. More generally, framework accommodate wider range structures—including full disparity between different levels fidelity—than existing kernel approaches based Δ-learning although show two be similar. Consequently, tool for reducing costs even further opportunistically exploiting