Neural Constrained Charge Equilibration for Extracting Charge Transport Parameters from Machine Learned Interatomic Potentials

Machine-learned interatomic potentials (MLIPs) continue to gain popularity as efficient tools for predicting energies and forces for atomic structures at an accuracy comparable to quantum chemistry. However, there are limited trained MLIPs for certain important applications such as in charge transport. Accelerating and automating the determination of parameters for such applications can be of critical importance in the rapid discovery of materials such as organic electronics (e.g., OLEDs). This work develops and introduces neural constrained charge equilibriation (NcQEq) and applies the technique to AimNet2 and OrbMol-V2. The new method enables extracting information about the diabatic states and thereby charge transport parameters. The approach is demonstrated on model systems against constrained density functional theory charge transport and then tested on intramolecular hole transport properties for polyethylene. This work demonstrates that the NcQeq approach is compatible with the charge equilibration schemes for AimNet2 and the latent charge predictions of OrbMol-V2. After calibration to reference data, the method enables estimation of the electronic coupling, reorganization energy, and the free energy difference.