ShiftML

Official Resources

• Source Repository: https://github.com/lab-cosmo/ShiftML
• Documentation: https://lab-cosmo.github.io/ShiftML/
• License: Open source (MIT-compatible)

Overview

ShiftML is a Python package for the prediction of NMR chemical shieldings in organic solids using machine learning. It is trained on a large dataset of chemical shieldings computed with DFT GIPAW and can predict shieldings for a wide range of organic crystals at a fraction of the computational cost.

Scientific domain: NMR chemical shift prediction, ML-accelerated spectroscopy
Target user community: Researchers predicting NMR chemical shifts for organic solids, NMR crystallography practitioners

Theoretical Methods

• Machine learning (Gaussian process regression)
• DFT GIPAW training data (CASTEP)
• SOAP (Smooth Overlap of Atomic Positions) descriptors
• Chemical shielding tensor prediction
• Isotropic and anisotropic shieldings
• NMR crystallography

Capabilities (CRITICAL)

• NMR chemical shielding prediction
• 1H, 13C, 15N, 17O, and other nuclei
• Isotropic shielding
• Shielding anisotropy
• Fast ML prediction (milliseconds per site)
• Organic crystal support
• Structure-to-shielding mapping
• Uncertainty estimation (Gaussian process)

Sources: GitHub repository, Chem. Sci. 12, 3365 (2021)

Key Strengths

ML Acceleration:

• Orders of magnitude faster than DFT GIPAW
• Milliseconds per shielding prediction
• Enables high-throughput NMR prediction
• Uncertainty quantification

DFT-Quality Accuracy:

• Trained on CASTEP GIPAW data
• 1.4 million shieldings in training set
• Systematic improvement with more data
• Validated against experiment

Organic Solids Focus:

• Optimized for organic crystals
• Handles disorder and polymorphism
• Supports common organic elements
• NMR crystallography applications

Uncertainty Estimation:

• Gaussian process regression
• Meaningful error bars
• Active learning possible
• Confidence in predictions

Inputs & Outputs

• Input formats:

  • Atomic structures (ASE, CIF, XYZ)
  • Crystal structure files

• Output data types:

  • Chemical shieldings (isotropic)
  • Shielding anisotropy
  • Prediction uncertainties
  • Site-resolved shieldings

Interfaces & Ecosystem

• ASE: Structure handling
• CASTEP: Training data source
• QUIP/SOAP: Descriptors
• Python/NumPy: Computation

Performance Characteristics

• Speed: Milliseconds per site
• Accuracy: ~1-2 ppm for 13C, ~0.3 ppm for 1H
• System size: Any organic crystal
• Memory: Low (trained model)

Computational Cost

• Prediction: Milliseconds per site
• Training: Pre-computed (done once)
• vs GIPAW: 1000x+ speedup
• Typical: Negligible

Limitations & Known Constraints

• Organic solids only: Not for inorganic/metallic systems
• Training domain: Limited to organic crystal compositions
• No dynamics: Static structure only
• GIPAW reference: Accuracy limited by GIPAW quality
• Heavy elements: Limited support

Comparison with Other Codes

• vs CASTEP GIPAW: ShiftML is ML-fast, CASTEP is DFT-accurate
• vs MRSimulator: ShiftML predicts shieldings, MRSimulator simulates spectra
• vs ml4nmr: ShiftML is for solids, ml4nmr corrects molecular DFT
• Unique strength: ML prediction of NMR chemical shieldings for organic solids, uncertainty estimation

Application Areas

NMR Crystallography:

• Structure validation via NMR
• Polymorph identification
• Disorder characterization
• Crystal structure determination

Pharmaceutical Materials:

• API polymorph screening
• Cocrystal characterization
• Amorphous solid dispersions
• Hydrate/anhydrate identification

Organic Semiconductors:

• Molecular packing determination
• π-π stacking characterization
• Charge transport correlations
• Structure-property relationships

Metal-Organic Frameworks:

• Linker conformation analysis
• Guest molecule identification
• Framework dynamics
• Functional group characterization

Best Practices

Structure Preparation:

• Use well-optimized structures
• Include hydrogen positions
• Consider disorder models
• Validate geometry

Prediction Analysis:

• Use uncertainty estimates
• Compare with experimental shifts
• Account for temperature effects
• Consider dynamics averaging

Community and Support

• Open source on GitHub
• Developed at COSMO Lab (EPFL)
• Published in Chem. Sci.
• Active development
• Documentation available

Verification & Sources

Primary sources:

1. GitHub repository: https://github.com/lab-cosmo/ShiftML
2. Documentation: https://lab-cosmo.github.io/ShiftML/
3. M. Caro et al., Chem. Sci. 12, 3365 (2021)

Confidence: VERIFIED

Verification status: ✅ VERIFIED

• Source code: ACCESSIBLE (GitHub)
• Documentation: ACCESSIBLE
• Published methodology: Chem. Sci.
• Active development: Ongoing
• Specialized strength: ML prediction of NMR chemical shieldings for organic solids, uncertainty estimation