Physics-Informed and Equivariant Machine Learning for Molecular Dipole Moment Prediction

Accurate prediction of molecular electric dipole moments is crucial for determining molecular properties and understanding molecular reactivity. Recent developments in equivariant machine learning enable direct prediction of electric dipole moments as vector quantities. This raises the question as to whether it is still necessary to consider physics-informed charge-based principles in the model to achieve satisfactory performance. In this work, we systematically compare direct equivariant electric dipole prediction using MACE with charge-based approaches using adapted MACE which incorporates a variant charge equilibrium equation (QEq) into the MACE framework. Across diverse datasets including QM7b, QM9, as well as subsets of the SPICE and SN2 data sets, we demonstrate that both direct equivariant electric dipole prediction and physics-informed QEq approaches show good performance for short-to-medium range interaction datasets. The charge-based QEq model shows slightly better performance than direct prediction beyond small training data sizes. Notably, the charge-based QEq model outperforms direct electric dipole prediction for the systems with long range interactions. Our results highlight the importance of incorporating physics into the model for improved model interpretability and transferability.