Induced Dipole Calculation with E(3)-Equivariant Neural Networks and Multipole Field Perturbation.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Polarizable force fields are crucial for molecular dynamics (MD) simulations of biological systems, capturing electric induction effects via induced dipoles.
• Iterative computation of induced dipoles presents convergence challenges, particularly in large-scale simulations.

Purpose of the Study:

• To implement an E(3)-equivariant neural network for predicting induced dipoles, thereby avoiding iterative computations.
• To enhance the efficiency and accuracy of molecular dynamics simulations for polar solvent systems.

Main Methods:

• Developed an E(3)-equivariant neural network architecture to predict induced dipoles.
• Integrated a physics-informed loss function to utilize artificially perturbed training data.
• Validated the network on water systems and benchmarked performance across varying densities, system sizes, and ice polymorphs.

Main Results:

• The neural network successfully predicted induced dipoles, circumventing iterative convergence issues.
• Perturbation-based data augmentation significantly improved model transferability across diverse chemical environments.
• Physics-informed loss alone demonstrated limited generalization benefits.

Conclusions:

• E(3)-equivariant neural networks offer a viable, iteration-free alternative for calculating induced dipoles in polar systems.
• Data augmentation strategies are key to enhancing the robustness and applicability of these models in molecular dynamics.
• The developed method shows promise for accelerating simulations of complex biological and material systems.