Machine-Learning X-Ray Absorption Spectra to Quantitative Accuracy.

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Simulations of excited state properties, like spectral functions, are computationally intensive.
• This limits their application in high-throughput materials modeling and discovery.

Purpose of the Study:

• To demonstrate the efficacy of graph-based neural networks for predicting molecular x-ray absorption near-edge structure (XANES) spectra.
• To establish a computationally efficient method for spectral analysis and structure inference.

Main Methods:

• Utilized graph-based neural networks for spectral property prediction.
• Trained and validated the model on molecular XANES spectra.
• Quantitatively assessed prediction accuracy against ground truth data.

Main Results:

• Achieved quantitative accuracy in predicting XANES spectra.
• Reproduced nearly all prominent spectral peaks.
• 90% of predicted peak locations were within 1 eV of experimental or computed values.

Conclusions:

• Graph neural networks offer a viable and accurate alternative to computationally expensive simulations for XANES spectra.
• This approach facilitates high-throughput spectral sampling, accelerating materials design and discovery.
• The method holds potential for spectral analysis and molecular structure inference.