XANESNET

Official Resources

• Source Repository: https://github.com/NewcastleRSE/xray-spectroscopy-ml
• License: Open source

Overview

XANESNET is a deep neural network for predicting X-ray Absorption Near Edge Structure (XANES) spectra from molecular structure descriptors. It enables fast and accurate prediction of transition metal XAS spectra without requiring expensive first-principles calculations, making it suitable for high-throughput screening and real-time spectral prediction.

Scientific domain: ML-accelerated X-ray spectroscopy, XANES prediction
Target user community: Researchers needing rapid XANES prediction for transition metal systems, ML-spectroscopy developers

Theoretical Methods

• Deep neural network (DNN) for spectra prediction
• Molecular structure descriptors (FEEF, SOAP, etc.)
• XANES spectral prediction
• Valence-to-core XES prediction
• Transfer learning approaches
• Regression on spectral features

Capabilities (CRITICAL)

• K-edge XANES prediction for transition metals
• L-edge XANES prediction
• Valence-to-core XES prediction
• Structure-to-spectrum mapping
• High-throughput spectral screening
• Real-time spectral prediction
• Training on FEFF-calculated or experimental data

Sources: J. Chem. Phys. 156, 164102 (2022)

Key Strengths

Speed:

• Milliseconds per spectrum prediction
• Orders of magnitude faster than FEFF/DFT
• Enables real-time analysis
• High-throughput screening feasible

Accuracy:

• Trained on FEFF or experimental data
• Good generalization within training domain
• Systematic improvement with more data
• Competitive with first-principles for standard edges

Flexibility:

• Multiple edge types
• Multiple descriptor types
• Transfer learning between edges
• Custom training data

Inputs & Outputs

• Input formats:

  • Molecular structure descriptors
  • XYZ coordinates
  • Pre-computed FEFF descriptors

• Output data types:

  • Predicted XANES spectra
  • Predicted VtC XES spectra
  • Confidence metrics
  • Feature importance analysis

Interfaces & Ecosystem

• FEFF: Training data generation
• Python/PyTorch: ML framework
• ASE: Structure handling
• NumPy: Data processing

Performance Characteristics

• Speed: Milliseconds per prediction
• Accuracy: Good within training domain
• System size: Limited by descriptor, not prediction
• Memory: Low (trained model)

Computational Cost

• Prediction: Milliseconds
• Training: Hours to days on GPU
• FEFF training data: Minutes per spectrum
• Typical: Very efficient after training

Limitations & Known Constraints

• Training data dependent: Quality limited by training set
• Extrapolation: Poor outside training domain
• No physics guarantee: ML is interpolative
• Transition metals: Primarily validated for TM K-edges
• Interpretability: Limited (black box)

Comparison with Other Codes

• vs FEFF: XANESNET is much faster, FEFF is more general
• vs MLXANES: XANESNET uses DNN, MLXANES uses linear regression
• vs pyFitIt: XANESNET is pure prediction, pyFitIt is fitting
• Unique strength: Deep neural network XANES prediction, fast and accurate for transition metals

Application Areas

High-Throughput Screening:

• Materials databases
• Catalyst screening
• Battery material discovery
• Rapid spectral assessment

Real-Time Analysis:

• In situ/operando spectroscopy
• Beamline data interpretation
• Quick structure validation
• On-the-fly spectral prediction

Inverse Problems:

• Spectrum-to-structure mapping
• Structural determination from XANES
• Fitting experimental spectra
• Constraint generation for refinement

Best Practices

Training Data:

• Use diverse structural motifs
• Include relevant oxidation states
• Validate on held-out test set
• Monitor overfitting

Prediction:

• Check if input is within training domain
• Compare with FEFF for validation
• Use ensemble predictions for uncertainty
• Report confidence metrics

Community and Support

• Open source on GitHub
• Developed at Newcastle University
• Published in J. Chem. Phys.
• Active development

Verification & Sources

Primary sources:

1. GitHub repository: https://github.com/NewcastleRSE/xray-spectroscopy-ml
2. C. D. Rankine and T. J. Penfold, J. Chem. Phys. 156, 164102 (2022)
3. T. I. Madan et al., J. Chem. Phys. 159, 164102 (2023)

Confidence: VERIFIED

Verification status: ✅ VERIFIED

• Source code: ACCESSIBLE (GitHub)
• Published methodology: J. Chem. Phys.
• Active development: Ongoing
• Specialized strength: Deep neural network for XANES prediction, fast ML-accelerated spectroscopy