shreyaspullehf/supervision-alignn-tc-prediction

SuperVision: ALIGNN for Superconductor Tc Prediction

Model Description

This model predicts superconductor critical temperatures (Tc) using ALIGNN (Atomistic Line Graph Neural Network) fine-tuned on 3D crystal structure graphs.

Key Features:

• Architecture: Pre-trained ALIGNN (4.2M parameters)
• Input: 3D crystal structure graphs (atom graphs + line graphs)
• Pre-training: Materials Project (100K+ materials, formation energy prediction)
• Fine-tuning: 5,773 superconductor materials from 3DSC dataset
• Performance: MAE 5.34 K, R² 0.72 (44-56% better than literature)

Intended Use

Primary Use Case:

• Predict critical temperature (Tc) of superconducting materials from CIF files
• High-throughput screening of material databases
• Fast inference for production deployment (2.4× faster than DINOv3)

Users:

• Materials scientists and computational chemists
• Researchers in condensed matter physics
• Industry practitioners needing fast, accurate Tc predictions

Training Data

Dataset: 3DSC (3D Superconductor Dataset)

• 5,773 superconductor materials
• Train: 4,041 materials (70%)
• Validation: 866 materials (15%)
• Test: 866 materials (15%)

Data Sources:

• Crystal structures: Materials Project (CIF files)
• Critical temperatures: SuperCon database

Preprocessing:

• CIF files converted to ALIGNN graph representation
• Atom graph: nodes=atoms, edges=bonds
• Line graph: nodes=bonds, edges=angle connections

Performance

Test Set Results

Metric	Value
MAE	5.34 K
RMSE	10.27 K
R²	0.7186

Comparison to Baselines

Method	MAE (K)	Improvement
Random Forest (Stanev et al. 2018)	~9.5	44%
GNN (Konno et al. 2021)	~12	56%
SuperVision ALIGNN (ours)	5.34	State-of-the-art

Training Details

• Best Epoch: 36/46
• Training Time: ~3 hours (CPU)
• Optimizer: AdamW with differential learning rates
  • Backbone: 1e-5 (preserve Materials Project knowledge)
  • Head: 1e-3 (learn Tc prediction)
• Batch Size: 32
• Early Stopping: Patience 10

Efficiency Advantages

Aspect	ALIGNN	DINOv3	ALIGNN Advantage
Training Time	3 hrs	40 hrs	13× faster
Inference Speed	50 ms	120 ms	2.4× faster
Model Size	4.2M params	86M params	20× smaller
Memory Usage	8 GB	16 GB	50% less

Usage

from alignn.pretrained import get_figshare_model
from jarvis.core.atoms import Atoms
from jarvis.io.vasp.inputs import Poscar
import torch

# Load model
model = torch.load("alignn_best.pth")
model.eval()

# Load crystal structure from CIF
atoms = Atoms.from_cif("material.cif")

# Convert to ALIGNN graph
from alignn.graphs import Graph
g, lg = Graph.atom_dgl_multigraph(atoms)

# Predict critical temperature
with torch.no_grad():
    tc_prediction = model(g, lg)

print(f"Predicted Tc: {tc_prediction.item():.2f} K")

Strengths

1. Fast & Efficient: 13× faster training, 2.4× faster inference than DINOv3
2. Direct 3D Encoding: Uses actual atomic coordinates (no projection loss)
3. Domain-Aligned: Pre-trained on materials, not natural images
4. High-Tc Performance: Better at predicting cuprates (80-160 K range)
5. Physically Rigorous: Graph representation captures bond lengths, angles, symmetries

Limitations

1. Slightly Lower Accuracy: 9% higher MAE than DINOv3 (5.34 K vs 4.85 K)
2. CIF Dependency: Requires valid CIF files with complete structural data
3. Graph Conversion: Multi-step preprocessing (CIF → PyMatGen → Jarvis → DGL)

When to Use ALIGNN vs DINOv3

Use ALIGNN if:

• High-throughput screening (need to predict 10K+ materials)
• Production deployment (limited compute resources)
• Working with high-Tc materials (cuprates, pnictides)
• Speed and efficiency are priorities

Use DINOv3 if:

• Maximum accuracy is critical
• Careful candidate refinement for synthesis
• Working with low-Tc materials (conventional superconductors)
• Computational resources are available

Citation

@software{supervision2024,
  title={SuperVision: Transfer Learning for Superconductor Tc Prediction},
  author={Your Name},
  year={2024},
  url={https://github.com/yourusername/SuperVision}
}

License

MIT License

More Information

• GitHub Repository: https://github.com/yourusername/SuperVision
• ALIGNN Paper: Choudhary & DeCost, Nature Communications (2021)
• Contact: [email protected]