Ab initio parametrization of distributed polarizable force fields
Polarizable force fields offer superior transferability and accuracy compared to classical force fields, enabling access to electronic response properties such as refractive index and electronic density of states. Here, we demonstrate two key improvements that significantly enhance their accuracy: (1) assigning atomic polarizability to individual atoms rather than atom types, and (2) employing atomic polarizability tensors instead of scalar values. These modifications extend the applicability of polarizable force fields to cations, anions, and excited states, while also providing more accurate descriptions of neutral molecules. We propose a first-principles-based parameterization procedure for atomic polarizability tensors and scalars, validated on a set of small organic molecules with conjugated building blocks. To overcome the computational cost of ab initio calculations, we train a message-passing graph neural network to predict polarizability parameters, enabling efficient and scalable parameterization. Crucially, this approach imposes no additional computational cost during simulations and provides a clear diagnostic criterion for identifying cases where polarizable force field models fail to accurately describe molecular polarizability.