Digital Alchemy: ML for Reaction Yield Prediction

Machine learning for predicting Buchwald-Hartwig C-N coupling reaction yields. Compares one-hot encoding vs transformer-based approaches (ChemBERTa) on the Ahneman dataset (4312 reactions).

Course Project: Developed as part of an ML course, including chemistry sessions and practical exercises.

---

Results

Approach	RMSE	MAE	R²
One-Hot + Neural Network	7.8%	5.6%	0.9
One-Hot + Gradient Boosting	11.1%	8.3%	0.8
ChemBERTa + Neural Network	16.2%	11.5%	0.6

Finding: One-hot encoding significantly outperforms transformers for fixed-component datasets (46 unique molecules).

---

Quick Start

Requirements: Python 3.11

# Create virtual environment with Python 3.11
python3.11 -m venv alchemy_env
source alchemy_env/bin/activate  # Linux/Mac
# alchemy_env\Scripts\activate  # Windows

# Install all dependencies
pip install -r requirements.txt

# Run web app
python yield_predictor_Gradio.py

Web Interface: Select catalyst, aryl halide, base, and additive → Get instant yield prediction

Note: Python 3.11 is required for proper dependency resolution between TensorFlow and ord-schema.

---

Repository

• yield_prediction_ahneman.ipynb: One-hot encoding approach
• yield_prediction_Reaction_Smiles.ipynb: ChemBERTa approach
• yield_predictor_Gradio.py: Interactive web app
• yield_model.keras: Trained model
• requirements.txt: Python dependencies
• Final_Report.pdf: Full technical report

---

Dataset: Ahneman Buchwald-Hartwig from Open Reaction Database