Molecular dynamics with on-the-fly machine learning of quantum-mechanical forces

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Molecular dynamics simulations are crucial for understanding material properties.
• First-principles quantum mechanical (QM) calculations offer high accuracy but are computationally expensive.
• Machine learning (ML) models can accelerate simulations but typically require large, complete datasets.

Purpose of the Study:

• To develop an information-efficient molecular dynamics scheme.
• To integrate first-principles QM and ML techniques seamlessly.
• To reduce the computational cost of accurate molecular dynamics simulations.

Main Methods:

• A hybrid approach combining Bayesian inference for ML force predictions and on-the-fly QM calculations.
• A growing ML database that incorporates QM data as needed.
• Progressive learning where fewer QM calculations are required for repeated processes.

Main Results:

• Demonstrated accuracy and generality of the combined QM/ML scheme.
• Significant reduction in QM calls for encountered chemical processes.
• Successful application to crystalline and molten silicon systems.

Conclusions:

• The proposed scheme offers an accurate and efficient alternative to traditional methods.
• The adaptive nature of the ML database enhances computational efficiency.
• This approach paves the way for simulating complex chemical processes with reduced computational burden.