Fast, modular, and differentiable framework for machine learning-enhanced molecular simulations.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Molecular dynamics and Monte Carlo simulations are crucial for understanding material properties.
• Existing frameworks often lack flexibility or computational efficiency.
• Integrating machine learning potentials with classical methods presents a challenge.

Purpose of the Study:

• To introduce DIMOS, an end-to-end differentiable molecular simulation framework.
• To enable seamless integration of machine learning potentials and classical force fields.
• To achieve significant performance improvements in molecular simulations.

Main Methods:

• Developed a modular, differentiable framework supporting molecular dynamics and Monte Carlo methods.
• Implemented efficient classical force fields and machine learning potential integration.
• Utilized PyTorch for flexibility and incorporated optimized algorithms for scalability.

Main Results:

• DIMOS achieves up to 170x speed-up for classical force field simulations compared to other differentiable frameworks.
• Demonstrated end-to-end optimization of simulation parameters, leading to a 3x acceleration.
• Showcased linear scaling with system size, overcoming quadratic scaling limitations.

Conclusions:

• DIMOS provides a flexible and highly efficient platform for molecular simulations.
• The framework bridges the gap between traditional simulation engines and modern machine learning approaches.
• Differentiability offers a powerful tool for optimizing simulation parameters and improving accuracy.