Differentiable Optimization for the Prediction of Ground State Structures (DOGSS).

Area of Science:

• Computational Materials Science
• Surface Science
• Solid State Chemistry

Background:

• Ground state structures are crucial for computational materials and surface science.
• Current methods lack general approaches to improve initial structure guesses from past calculations.

Purpose of the Study:

• To develop a general method for predicting ground state configurations.
• To improve initial structure guesses for inorganic materials and surface adsorbates.

Main Methods:

• Utilized differentiable optimization and graph neural networks.
• Learned properties of a simple harmonic force field to approximate ground state structures.
• Applied the method to inorganic multicomponent surfaces and adsorbate relaxations.

Main Results:

• Demonstrated a flexible, open-source tool for improving initial configurations.
• Successfully applied to datasets across 32 elements for surface relaxations.
• Reduced computational cost of adsorbate-covered surface relaxations by approximately 50%.

Conclusions:

• The developed method effectively improves initial configurations for materials science calculations.
• This approach accelerates computationally expensive relaxation processes.
• It complements existing methods for computational materials discovery.