Scalable Boltzmann generators for equilibrium sampling of large-scale materials.

Area of Science:

• Computational chemistry and materials science
• Machine learning for scientific discovery
• Statistical mechanics and molecular modeling

Background:

• Traditional molecular dynamics simulations face challenges in sampling efficiency for large systems.
• Boltzmann Generators offered one-shot deep learning for equilibrium sampling but struggled with scalability.
• Developing efficient methods for generating equilibrium ensembles is crucial for molecular and materials modeling.

Purpose of the Study:

• To overcome the scalability limitations of previous Boltzmann Generator architectures for large materials systems.
• To develop a novel deep learning approach for efficient equilibrium structure ensemble generation.
• To enable accurate modeling of materials with large simulation cells.

Main Methods:

• Implemented a Boltzmann Generator architecture combining augmented coupling flows and graph neural networks.
• Utilized local environment exploitation for energy-based training and rapid inference.
• Applied the model to Lennard-Jones crystals, mW water ice phases, and the silicon phase diagram.

Main Results:

• Achieved faster training and reduced resource usage compared to previous designs.
• Demonstrated superior sampling efficiency for equilibrium structure generation.
• Successfully modeled large materials systems with over a thousand atoms, showing negligible finite-size effects.
• Produced accurate equilibrium ensembles and free energies across various scales.

Conclusions:

• The developed Boltzmann Generator architecture effectively scales to large materials systems.
• This approach significantly enhances sampling efficiency and accuracy in molecular and materials modeling.
• The method provides a powerful tool for studying materials properties and phase behavior at unprecedented scales.