Uncertainty Based Machine Learning-DFT Hybrid Framework for Accelerating Geometry Optimization.

Area of Science:

• Computational Chemistry
• Materials Science
• Physics

Background:

• Geometry optimization is crucial for computational simulations in chemistry, physics, and materials science.
• Current methods often require numerous force evaluations, limiting computational efficiency and the pace of materials discovery.

Purpose of the Study:

• To develop a more efficient and reliable algorithm for geometry optimization.
• To accelerate computational modeling and materials discovery through reduced force evaluations.

Main Methods:

• A delta method-based neural network-density functional theory (DFT) hybrid optimizer was developed.
• Key innovations include a modified delta method for enhanced uncertainty estimation and a quasi-Newton approach using a neural network-derived Hessian for improved stability.

Main Results:

• The hybrid optimizer was benchmarked against standard algorithms on various systems (bulk metal, metal surface, metal hydride, oxide cluster).
• The proposed optimizer consistently reduced the number of DFT force calls by 2-3 times across all tested systems.

Conclusions:

• The novel hybrid optimizer significantly enhances the computational efficiency of geometry optimization.
• This advancement holds promise for accelerating computational modeling and facilitating new materials discovery.