Enhancing transferability of machine learning-based polarizability models in condensed-phase systems via atomic polarizability constraint

Accurate prediction of molecular polarizability is essential for understanding electrical, optical, and dielectric properties of materials. Traditional quantum mechanical (QM) methods, though precise for small systems, are computationally prohibitive for large-scale systems. In this work, we proposed an efficient approach for calculating molecular polarizability of condensed-phase systems by embedding atomic polarizability constraints into the tensorial neuroevolution potential (TNEP) framework. Using n-heneicosane as a benchmark, a training data set was constructed from molecular clusters truncated from the bulk systems. Atomic polarizabilities derived from semi-empirical QM calculations were integrated as training constraints for its balance of computational efficiency and physical interpretability. The constrained TNEP model demonstrated improved accuracy in predicting molecular polarizabilities for larger clusters and condensed-phase systems, attributed to the model’s refined ability to properly partition molecular polarizabilities into atomic contributions across systems with diverse configurational features. Results highlight the potential of the TNEP model with atomic polarizability constraint as a generalizable strategy to enhance the scalability and transferability of other atom-centered machine learning-based polarizability models, offering a promising solution for simulating large-scale systems with high data efficiency.