Optimal Transport Distances to Characterize Electronic Excitations

Understanding the character of electronic excitations is important in computational and reaction mechanistic studies, but their classification from simulations remains an open problem. Distances based on optimal transport have proven very useful in a plethora of classification problems and, therefore, seem a natural tool to try to tackle this challenge. We propose and investigate a new diagnostic Theta based on the Sinkhorn divergence from optimal transport. We evaluate a k-NN classification algorithm on Theta, the popular Lambda diagnostic, and their combination, and assess their performance in labeling excitations, finding that (i) the combination only slightly improves the classification, (ii) Rydberg excitations are not separated well in any setting, and (iii) Theta breaks down for charge transfer in small molecules. We then define a length-scale-normalized version of Theta and show that the result correlates closely with Lambda for results obtained with Gaussian basis functions. Finally, we discuss the orbital dependence of our approach and explore an orbital-independent version. Using an optimized combination of the optimal transport and overlap diagnostics together with a different metric is in our opinion the most promising for future classification studies.