Ultrafast Electronic Coupling Estimators: Neural Networks versus Physics-Based Approaches.

Fast and accurate estimation of electronic coupling matrix elements between molecules is essential for the simulation of charge transfer phenomena in chemistry, materials science, and biology. Here we investigate neural-network-based coupling estimators combined with different protocols for sampling reference data (random, farthest point, and query by committee) and compare their performance to the physics-based analytic overlap method (AOM), introduced previously. We find that neural network approaches can give smaller errors than AOM, in particular smaller maximum errors, while they require an order of magnitude more reference data than AOM, typically one hundred to several hundred training points, down from several thousand required in previous ML works. A Δ-ML approach taking AOM as a baseline is found to give the best overall performance at a relatively small computational overhead of about a factor of 2. Highly flexible π-conjugated organic molecules like non-fullerene acceptors are found to be a particularly challenging case for ML because of the varying (de)localization of the frontier orbitals for different intramolecular geometries sampled along molecular dynamics trajectories. Here the local symmetry functions used in ML are insufficient, and long-range descriptors are expected to give improved performance.