🔬 Electrocatalyst Quality Prediction using Machine Learning

A comprehensive machine learning project that predicts electrocatalyst performance using Random Forest classification with advanced feature engineering and permutation feature importance analysis.

📋 Project Overview

This project develops a machine learning pipeline to classify electrocatalyst quality based on overpotential measurements at 50.0 mA/cm². By analyzing 500 electrochemical experiments with 11 compositional and operational features, the model predicts whether catalysts are "good" (|η| < 0.6 V) or "bad" (|η| ≥ 0.6 V).

Key Achievement: Built a Random Forest classifier achieving 70.6% precision on test data, enabling reliable catalyst quality screening and reducing the need for expensive experimental validation.

🎯 Business Impact

• Cost Reduction: Predicts catalyst quality before expensive electrochemical testing
• Accelerated R&D: Identifies key compositional factors (Co, Se, Ni) for catalyst optimization
• Data-Driven Design: Provides actionable insights for next-generation catalyst development

🛠️ Technologies & Skills Demonstrated

Core Technologies

• Python 3.12 | scikit-learn | pandas | NumPy
• Matplotlib | Seaborn | PCA | Random Forest

Machine Learning Techniques

• ✅ High-dimensional data visualization (PCA, correlation analysis)
• ✅ Hyperparameter tuning with stratified validation
• ✅ Permutation Feature Importance (PFI) analysis
• ✅ Model evaluation with imbalanced datasets
• ✅ Scale-invariance validation for tree-based models

Data Science Workflow

1. Exploratory Data Analysis → Visualized 11D feature space using PCA and correlation heatmaps
2. Feature Engineering → Identified top correlated features and non-linear patterns
3. Model Development → Trained and optimized Random Forest with 5 hyperparameter configurations
4. Model Interpretation → Validated feature importance with/without normalization
5. Performance Analysis → Evaluated precision-recall trade-offs for imbalanced classes

📊 Dataset

• Source: Electrochemical experiments dataset
• Size: 500 samples
• Features: 11 compositional and operational parameters
  • Compositional: V, Cr, Mg, Fe, Co, Ni, Cu, S, Se, P
  • Operational: Voltage, Time
• Target: Overpotential η at 50.0 mA/cm²
• Class Distribution: 33% good / 67% bad (moderate imbalance)

🚀 Key Results

Model Performance

Metric	Training	Validation	Test
Precision	1.000	0.667	0.706
Recall	1.000	0.480	0.500
F1 Score	1.000	0.558	0.585

Top 5 Most Important Features

1. Cobalt (Co) - Importance: 0.211 ± 0.079
2. Selenium (Se) - Importance: 0.118 ± 0.076
3. Nickel (Ni) - Importance: 0.078 ± 0.053
4. Vanadium (V) - Importance: 0.049 ± 0.062
5. Magnesium (Mg) - Importance: 0.044 ± 0.026

Insight: Transition metals (Co, Ni) and chalcogens (Se) dominate catalyst performance—aligning with electrochemistry domain knowledge.

Hyperparameter Optimization

Tested 5 Random Forest configurations, achieving best validation F1 (0.565) with:

• n_estimators=100
• max_depth=20
• min_samples_split=5
• min_samples_leaf=2

📁 Project Structure

.
├── homework1_analysis.ipynb       # Complete analysis pipeline (32 cells)
├── ExerciseData.csv                # Dataset (500 experiments)
├── Homework1_Documentation.md      # Detailed technical report
└── README.md                       # This file

🔍 Methodology Highlights

1. High-Dimensional Visualization

• PCA Analysis: 2D projection explains only 24.4% variance → True high-dimensionality
• Correlation Heatmap: Identified Co (-0.393), Se (+0.352), V (+0.316) as top correlates
• Pairplots: Visualized non-linear class separability across top features

2. Stratified Data Splitting

• 70% Training (350 samples) - Model learning
• 15% Validation (75 samples) - Hyperparameter selection
• 15% Test (75 samples) - Unbiased evaluation

3. Feature Importance Validation

Ran PFI analysis with and without normalization to validate Random Forest's scale-invariance:

• ✅ Identical top-5 rankings in both scenarios
• ✅ Confirms importance stems from information gain, not feature magnitude

💡 Key Insights

1. Model Selection Rationale: Random Forest outperforms linear models due to:

   • PCA's low explained variance (35% in 3D) → non-linear relationships dominate
   • Heavy class intermixing in PCA space → complex decision boundaries required

2. Precision-Recall Trade-off: Model prioritizes precision (0.71) over recall (0.50):

   • Conservative predictions minimize false positives
   • Suitable for screening applications where experimental validation follows

3. Domain Alignment: Top features (Co, Se, Ni) match electrochemistry literature:

   • Transition metals enable variable oxidation states for redox reactions
   • Chalcogens modulate electronic structure and conductivity

🔧 Setup & Usage

Prerequisites

pip install numpy pandas scikit-learn matplotlib seaborn jupyter

Run the Analysis

jupyter notebook homework1_analysis.ipynb

The notebook runs end-to-end in ~30 seconds on modern hardware.

📈 Visualizations

The project includes 6 comprehensive visualizations:

1. Overpotential Distribution - Histogram with good/bad threshold
2. Correlation Heatmap - Feature-target relationships
3. Pairwise Scatter Plots - Top 5 features colored by quality
4. 2D PCA Projection - Class separation in reduced space
5. 3D PCA Projection - Multi-dimensional class structure
6. PFI Comparison - Normalized vs. non-normalized feature importance

🎓 Learning Outcomes

This project demonstrates:

• ✅ End-to-end ML pipeline design and implementation
• ✅ Imbalanced data handling with stratified sampling and F1 optimization
• ✅ Model interpretability through feature importance and domain validation
• ✅ Scientific rigor via normalization studies and cross-validation
• ✅ Professional documentation and reproducible research practices

🔮 Future Enhancements

• Implement SHAP values for instance-level interpretability
• Test gradient boosting methods (XGBoost, LightGBM, CatBoost)
• Engineer interaction features (e.g., Co×Se, Ni×Voltage)
• Perform threshold optimization for precision-recall balance
• Deploy model as REST API for real-time predictions

👤 Author

Abel Saj

• GitHub: @abelsaj
• LinkedIn: Connect with me