Machine-Learned Charge Transfer Integrals for Multiscale Simulations in Organic Thin Films

Gaining insight into structure-property relations is a key factor for the development of organic electronics. We present a multiscale framework for charge carrier mobilities in organic thin films empowered by machine-learned charge transfer integrals. The choice of the molecular representation is crucial for accurate and sensitive predictions. Using pentacene thin films, we investigate kernel based algorithms and systematically compare representations ranging from system-specific geometric to Coulomb matrix features to predict absolute and logarithmic transfer integrals. We use the predicted transfer integrals to compute the mobility, including its anisotropy, and compare it to reference values. Best accuracies were obtained by models using the interaction part of the Coulomb matrix as a feature and the logarithm of the transfer integral as a target. We achieve R2 values of 0.97 for transfer integrals within an extensive range of 20 orders of magnitude and less than 27% error in the mobility. We show the transferability of the CIP feature for tetracene and DNTT with excellent prediction accuracies. Furthermore, we demonstrate that the interaction part of the CM successfully encodes the molecular identity and provides a highly sensitive ML framework. The presented framework opens the possibility for highly accurate mesoscopic transport simulations saving orders of magnitude in computational cost.