SWMCP: Seawater-Aware Multi-Task Catalyst Potential

🔬 Novel Contributions

1. Seawater Condition Encoding

• Novel: First ML model to explicitly encode seawater electrochemical conditions (pH, salinity, Cl⁻ concentration, temperature) as learnable features
• Standard OC models (OC20/OC22/OC25) completely ignore experimental conditions

2. Multi-Task Physics-Informed Prediction

• Novel: Joint prediction of adsorption energy + surface energy + band gap + corrosion resistance
• Physics loss enforces Sabatier principle and scaling relations during training

3. Corrosion Resistance Score

• Novel: First ML model to explicitly predict seawater corrosion resistance for oxide electrocatalysts
• Critical for chloride-rich environments

4. Scaling Relation Enforcement

• Novel: Physics loss enforces dG_OOH ≈ dG_OH + 3.2 eV constraint
• Improves generalization to unseen catalyst compositions

🏗️ Architecture

Input: Crystal structure + Seawater conditions
  ↓
Seawater Condition Encoder (learnable pH, salinity, Cl⁻, temp)
  ↓
Equivariant Graph Neural Network (SE(3)-aware message passing)
  ↓
Multi-Task Prediction Heads:
  ├── Adsorption Energies (ΔG_OH, ΔG_O, ΔG_OOH)
  ├── Surface Energy (stability)
  ├── Band Gap (conductivity)
  ├── Corrosion Score (seawater-specific)
  └── Forces (per-atom)
  ↓
Physics-Informed Loss (Sabatier + Scaling + Stability)

📦 Files

• model.py - SWMCP model implementation
• train.py - Training script with dummy data loader

🚀 Usage

📊 Dataset Strategy

1. Pre-train on OC22 oxide subset (contains Bi, Cu, Cr separately)
2. Fine-tune on Bi-Cu, Bi-Cr, Cu-Cr binary oxides
3. Predict on Bi₂Cu₂Cr₂O₈ (zero-shot or with minimal DFT)

📄 Citation

• Open Catalyst Project (OC22, OC25)
• EquiformerV2/V3 papers for equivariant transformer architecture