⚗️ ReactionNet - Chemical Reactivity Prediction with Deep Learning

🇪🇸 Resumen: Red Neuronal Siamesa diseñada para predecir la reactividad química entre pares de elementos. Logra más del 96% de precisión utilizando embeddings de propiedades atómicas.

🇬🇧 English Profile: Siamese Neural Network architecture designed to predict chemical reactivity between element pairs. Achieves over 96% accuracy using atomic property embeddings.

DeepLearning models to predict if two chemical elements will react with each other.

🎯 Project Overview

ReactionNet uses Siamese Neural Networks to predict chemical reactivity between element pairs. The project includes:

• ✅ Two trained models (V1 and V2) with 96%+ accuracy
• ✅ 10,000 reaction pairs from real scientific literature
• ✅ Interactive web demo
• ✅ Complete training notebooks for Google Colab
• ✅ Both PyTorch and ONNX formats

📊 Model Performance

Model	Accuracy	F1-Score	AUC-ROC	Parameters
V1 (Recommended)	96.07%	96.11%	0.9950	39K
V2 (Transfer Learning)	95.67%	95.68%	0.9940	150K

Winner: ReactionNet V1 - simpler, faster, and more accurate! ✨

🚀 Quick Start

Option 1: Try the Web Demo

Open demo/index.html in your browser for an instant interactive experience.

Option 2: Use Pre-trained Models

Python with PyTorch:

import torch

# Load model
model = ReactionNetV1()
model.load_state_dict(torch.load('models/reactionnet_v1.pth'))
model.eval()

# Predict (example: Na + Cl)
elem1 = torch.tensor([[0.085, 0.077, 0.233, 0.031, 0.206, 0.056, 0.286, 0.0]])
elem2 = torch.tensor([[0.131, 0.118, 0.790, 0.117, 0.531, 0.944, 0.429, 0.333]])

probability = model(elem1, elem2)
print(f"Reaction probability: {probability.item():.4f}")

Python with ONNX:

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession('models/reactionnet_v1.onnx')
result = session.run(None, {'element_1': elem1, 'element_2': elem2})
print(f"Probability: {result[0][0][0]:.4f}")

Option 3: Train Your Own

1. Open notebooks/ReactionNet_V1_Training.ipynb in Google Colab
2. Upload data/reaction_pairs_lite.json
3. Run all cells
4. Download your trained model

📁 Repository Structure

ReactionNet_GitHub_Package/
├── models/               # Trained models (V1 and V2)
│   ├── reactionnet_v1.pth
│   ├── reactionnet_v1.onnx
│   ├── reactionnet_v2.pth
│   └── reactionnet_v2.onnx
│
├── data/                 # Training datasets
│   ├── reaction_pairs_lite.json (10K pairs)
│   ├── inorganic_reactions_curated.json (5K reactions)
│   └── periodic_table_master.json (130 elements)
│
├── results/              # Training reports and metrics
│   ├── v1/              # V1 results
│   └── v2/              # V2 results
│
├── demo/                 # Interactive web demo
│   └── index.html
│
├── notebooks/            # Training notebooks
│   ├── ReactionNet_V1_Training.py
│   └── ReactionNet_V2_Training.py
│
└── docs/                 # Documentation
    ├── MODEL_COMPARISON.md
    ├── TRAINING_GUIDE.md
    └── API_REFERENCE.md

🔬 Scientific Approach

Input Features (8 per element)

1. Atomic number (normalized)
2. Atomic mass (normalized)
3. Pauling electronegativity
4. Allen electronegativity
5. First ionization energy
6. Group number
7. Period number
8. Block (s/p/d/f)

Model Architecture

• Type: Siamese Neural Network with Attention
• Encoder: 8 → 128 → 64
• Attention: Learnable weights
• Fusion: Concatenate → 128 → 64 → 32 → 1
• Output: Sigmoid (reaction probability)

Training Details

• Dataset: 10,000 balanced pairs (50% positive, 50% negative)
• Source: Real reactions from CederGroup database
• Optimizer: AdamW with weight decay
• Scheduler: ReduceLROnPlateau
• Early Stopping: 25 epochs patience

📖 Documentation

• Model Comparison - V1 vs V2 detailed analysis
• Training Guide - How to train your own models
• API Reference - Complete API documentation

🎓 Citation

If you use this work in your research, please cite:

@software{speranza2026reactionnet,
  author = {Speranza, Maximiliano Rodrigo},
  title = {ReactionNet: Chemical Reactivity Prediction with Deep Learning},
  year = {2026},
  url = {https://github.com/SperanzaMax/reactionnet}
}

📈 Future Work

• Expand to organic reactions
• Predict specific reaction products
• Add temperature/pressure conditions
• Multi-reactant support (3+ elements)
• Reaction mechanism prediction

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

MIT License - see LICENSE file for details

👤 Author

Maximiliano Rodrigo Speranza
Email: maximiliano.speranza@gmail.com
Date: January 2026

---

⭐ If you find this useful, please star the repo!