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

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Mechanics

Background:

• Accurate prediction of molecular electric dipole moments is essential for understanding molecular behavior and reactivity.
• Equivariant machine learning models offer direct vector prediction of electric dipole moments.
• The necessity of integrating physics-based principles, like charge equilibrium, into these models remains an open question.

Purpose of the Study:

• To systematically compare direct equivariant electric dipole moment prediction with physics-informed, charge-based approaches.
• To evaluate model performance across diverse chemical datasets, including varying interaction ranges.
• To assess the impact of physics integration on model interpretability and transferability.

Main Methods:

• Employed the MACE (Machine-learning Atomic力) framework for both direct equivariant prediction and a charge-based approach.
• Integrated a variant of the charge equilibrium equation (QEq) into the MACE framework for physics-informed modeling.
• Tested models on QM7b, QM9, SPICE, and SN2 datasets, covering short-to-long range interactions.

Main Results:

• Both direct equivariant prediction and the physics-informed QEq approach demonstrated good performance on short-to-medium range interaction datasets.
• The charge-based QEq model exhibited slightly superior performance compared to direct prediction, particularly with larger training datasets.
• The QEq model significantly outperformed direct prediction for systems involving long-range interactions.

Conclusions:

• Physics-informed machine learning models, specifically the QEq approach, offer advantages in predicting molecular electric dipole moments.
• The integration of physics enhances model performance, especially for complex systems with long-range interactions.
• Incorporating physics is crucial for improving model interpretability and transferability in molecular property prediction.