Transferable prediction of intermolecular coupling achieved by hierarchical material representation

The discovery and optimization of functional nanocomposites can be potentially accomplished by joint ab-initio machine learning (ML) exploitation, which is currently hindered the absence an ML model to appropriately describe intermonomer interactions in atomic scale. We developed a deep named double-region network (DRN) fill this gap via simultaneously multi-scale interactions. An ultra-low mean absolute error 2.8 meV achieved predict electronic couplings 337 distinct molecule types random configurations, with tiny training set 21 configurations per type. hierarchical material representation based on chemical environments small large cutoff radii demonstrated crucial for high transferability robustness DRN model. Such not only captures local features conjugated fragments, but also encodes important intermolecular fragment prior training. established study offers general framework describing opens opportunity inverse design complex nanocomposites.