luokai0 Materials Informatics Expert
Before Starting
- Which material class? (metals, ceramics, polymers, 2D materials)
- Property prediction or structure generation?
- Computational or experimental data source?
Core Expertise Areas
Materials Databases
Materials Project: DFT-calculated properties for over 150,000 inorganic compounds. AFLOW: high-throughput DFT database with thermodynamic and electronic properties. OQMD: open quantum materials database, formation energies and stability. NOMAD: repository for experimental and computational materials data. Cambridge Structural Database: crystal structures from X-ray diffraction.
Featurization
Composition features: element fractions, stoichiometry-based descriptors. Structure features: radial distribution function, coordination environment. Electronic structure: band gap, DOS features from DFT calculations. SOAP: smooth overlap of atomic positions, invariant descriptor for local environments. Graph neural networks: represent crystal structure as graph, learn from topology.
ML for Property Prediction
CGCNN: crystal graph convolutional neural network, predict formation energy. MEGNet: graph network for molecules and crystals, multi-property prediction. Gaussian process: uncertainty quantification, active learning integration. Transfer learning: pre-train on large DFT dataset, fine-tune on small experimental.
High-Throughput Workflows
VASP and Quantum ESPRESSO: DFT codes for electronic structure calculations. pymatgen: Python library for materials analysis and database interface. Fireworks: workflow management for high-throughput calculations. Active learning: select most informative experiments to minimize label cost.
Best Practices
- Validate ML models on held-out data not used in training
- Use uncertainty quantification to guide experimental validation
- Check data quality and consistency before training
- Consider interpretability alongside prediction accuracy
Common Pitfalls
| Pitfall | Fix |
|---|---|
| Training and test set overlap | Ensure proper splitting by structure or composition |
| Ignoring data biases | Databases oversample certain chemistries |
| Overfitting small datasets | Use cross-validation and regularization |
| DFT-experiment gap | Account for systematic errors between DFT and experiment |
Related Skills
- machine-learning-expert
- computational-chemistry-expert
- physics/condensed-matter-expert