The machine learning force field trained with lots of quantum-level data may work.

Wen Yanbing Thank you for your response

Training a machine learning force field with extensive quantum-level data is a promising approach. Such models, when built with robust architectures like SchNet or PhysNet that respect physical symmetries and dynamic behavior, can potentially capture the fast fluctuations in dipole moments during molecular dynamics simulations. However, achieving quantum-level precision demands high-quality, diverse training data and effective uncertainty quantification to ensure reliability.