Artificial neural network for the configuration problem in solids.

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• Predicting material properties from atomic configurations is computationally intensive.
• Traditional methods require extensive simulations for each configuration.
• Efficiently exploring the configuration space is crucial for materials discovery.

Purpose of the Study:

• To apply artificial neural networks (ANN) for solving the configuration problem in solids.
• To establish a direct mapping from configuration vectors to energies.
• To reduce the computational burden in materials science research.

Main Methods:

• Utilizing an artificial neural network (ANN) model.
• Training the ANN with configuration vectors and corresponding energies.
• Benchmarking the ANN approach on the M1 phase of Mo-V-Te-Nb oxide.
• Incorporating geometry relaxation effects into the ANN training.

Main Results:

• The ANN approach significantly decreased computational burden by a factor of 20-50.
• Achieved high accuracy with R² = 0.96 and Mean Absolute Deviation (MAD) = 0.12 eV for configuration energies.
• Demonstrated the capability of ANN to handle geometry relaxation effects, yielding R² = 0.95 and MAD = 0.13 eV.

Conclusions:

• ANN provides an efficient and accurate method for predicting energies of solid configurations.
• The proposed machine learning approach substantially reduces computational cost in materials science.
• ANN models can effectively account for geometry relaxation, further enhancing their applicability.