Coarse-Graining with Equivariant Neural Networks: A Path Toward Accurate and Data-Efficient Models

Area of Science:

• Computational chemistry
• Molecular dynamics
• Machine learning

Background:

• Machine learning, particularly deep learning, is increasingly used in coarse-grained (CG) molecular modeling.
• Neural networks offer high accuracy for CG force fields by capturing multibody effects.
• A major drawback is the substantial amount of simulation data required for training.

Purpose of the Study:

• To address the data-hunger issue in training neural network-based CG force fields.
• To investigate the efficacy of equivariant convolutional operations in reducing data requirements.
• To demonstrate a method for more efficient training of CG models.

Main Methods:

• Incorporation of equivariant convolutional operations into neural network architectures.
• Training and evaluation of CG models for water using varying data sizes.
• Comparison of models with and without equivariant operations.

Main Results:

• Equivariant convolutional operations significantly reduce the data needed for training CG models.
• Models utilizing these operations can achieve functionality with as little as one frame of reference data.
• Standard neural networks without these operations fail to train effectively with limited data.

Conclusions:

• Equivariant convolutional operations offer a viable solution to the data requirements of deep learning in CG molecular modeling.
• This approach enhances the practicality and efficiency of developing accurate CG force fields.
• The findings pave the way for more accessible and data-efficient machine learning in molecular simulations.