CSPML

Overview

CSPML (Crystal Structure Prediction with Machine Learning) is a template-based crystal structure prediction method using metric learning for element substitution. It predicts stable structures by selecting templates from known crystal structures.

Theoretical Basis

• Template-based structure prediction
• Metric learning for element substitution
• Crystal structure similarity
• Formation energy prediction
• Template selection optimization

Key Capabilities

• Element substitution-based CSP
• Metric learning for template selection
• No ab initio calculations needed
• Fast structure prediction
• Wide chemical space coverage

Sources: Comp. Mater. Sci. 211, 111496 (2022)

Key Strengths

Speed:

• No DFT required
• Fast prediction
• Large-scale screening

Methodology:

• Metric learning
• Template selection
• Element substitution

Coverage:

• Wide chemical space
• Multiple crystal systems
• General applicability

Inputs & Outputs

• Input formats: Chemical composition
• Output data types: Predicted structures, template matches

Interfaces & Ecosystem

• Python: TensorFlow-based
• Databases: Template database
• Validation: DFT codes

Workflow and Usage

1. Input chemical composition
2. Run metric learning model
3. Select best templates
4. Generate substituted structures
5. Validate with DFT (optional)

Performance Characteristics

• Very fast prediction
• 50-65% accuracy on crystal systems
• Scales to large datasets

Computational Cost

• Minimal (ML inference)
• No DFT required
• Fast screening

Best Practices

• Use diverse template database
• Validate top predictions
• Consider multiple templates
• Check structural validity

Limitations & Known Constraints

• Template-dependent
• May miss novel structures
• Accuracy varies by system

Application Areas

• High-throughput screening
• Initial structure guessing
• Materials discovery
• Database expansion

Comparison with Other Codes

• vs TCSP: Similar approach, different ML
• vs USPEX: CSPML faster, less accurate
• Unique strength: Fast template-based prediction, no DFT

Community and Support

• Open-source (GitHub)
• Published methodology
• Code available

Verification & Sources

Primary sources:

1. GitHub: https://github.com/Minoru938/CSPML
2. Publication: Comp. Mater. Sci. 211, 111496 (2022)

Confidence: VERIFIED

Verification status: ✅ VERIFIED

• Website: ACTIVE (GitHub)
• Source: OPEN
• Development: MAINTAINED
• Applications: Fast CSP, template-based prediction