Evidential deep learning for interatomic potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning interatomic potentials (MLIPs) are crucial for large-scale molecular simulations, offering ab initio accuracy.
• Active learning iteratively expands training datasets using uncertainty to identify out-of-distribution data.
• Current uncertainty quantification (UQ) methods for MLIPs face challenges with computational expense or prediction accuracy trade-offs.

Purpose of the Study:

• To develop a novel evidential deep learning framework for UQ in MLIPs.
• To achieve accurate UQ without compromising computational efficiency or prediction accuracy.
• To provide a robust and efficient alternative for UQ in molecular simulations.

Main Methods:

• An evidential deep learning framework is proposed for interatomic potentials.
• The framework incorporates a physics-inspired design.
• Uncertainty quantification is integrated directly into the deep learning model.

Main Results:

• The proposed method achieves UQ with minimal computational overhead.
• Prediction accuracy is maintained, outperforming existing UQ methods across diverse datasets.
• Demonstrated applications in exploring atomic configurations for water and universal potentials.

Conclusions:

• The evidential deep learning framework offers a computationally efficient and accurate UQ solution for MLIPs.
• This approach enhances the reliability of large-scale molecular simulations.
• The method shows significant potential for advancing molecular simulation and materials discovery.