🧪 Melting Point Prediction

Ensemble Machine Learning solution for predicting melting points of organic compounds using SMILES molecular descriptors. This repository contains my solution for the Kaggle Thermophysical Property: Melting Point competition.

📋 Competition Overview

Predicting the melting point of organic molecules is a long-standing challenge in chemistry and chemical engineering. Melting point is critical for drug design, material selection, and process safety, yet experimental measurements are often costly, time-consuming, or unavailable.

Goal

Build ML models that predict melting point (in Kelvin) for organic compounds given molecular descriptors.

Evaluation Metric

Mean Absolute Error (MAE) - Lower is better.

📊 Dataset Description

Split	Samples	Percentage
Train	2,662	80%
Test	666	20%
Total	3,328	100%

Files

• train.csv: Features (SMILES) + Target (Tm)
• test.csv: Features only, no target
• sample_submission.csv: Template with columns [id, Tm]

Columns

• id: Unique identifier
• SMILES: Molecular string representation
• Group 1..N: Descriptor features
• Tm: Melting point in Kelvin (train only)

🛠️ Solution Approach

Feature Engineering

Extensive molecular feature extraction using RDKit:

1. Basic Descriptors: MolWt, LogP, TPSA, HBond donors/acceptors, Rotatable bonds
2. Ring Features: Aromatic/Aliphatic/Saturated ring counts, Ring density
3. Charge Features: Gasteiger charges (mean, std, max, min)
4. Fragment Counts: Benzene, phenol, ester, ether, aldehyde, ketone, etc.
5. Morgan Fingerprints: 1024-bit with radius 3
6. MACCS Keys: 167-bit structural keys
7. Interaction Features: HBond capacity, flexibility, polarity index, etc.

Total Features: ~1,233 features after processing

Model Architecture

Ensemble of three gradient boosting models with optimized hyperparameters:

Model	n_estimators	learning_rate	max_depth
LightGBM	2000	0.02	10
XGBoost	1500	0.03	8
CatBoost	1500	0.03	8

Ensemble Strategy

• 5-Fold Cross Validation
• Optimal weight optimization using scipy.optimize
• Final weights: ~35% LightGBM, ~32% XGBoost, ~33% CatBoost

📈 Results

Cross-Validation Scores (MAE)

Model	CV MAE
LightGBM	28.70
XGBoost	28.57
CatBoost	28.75
Ensemble	28.16

🚀 Quick Start

Prerequisites

pip install rdkit pandas numpy scikit-learn lightgbm xgboost catboost optuna scipy joblib tqdm matplotlib seaborn

Run the Solution

# Clone the repository
git clone https://github.com/adityapawar327/melting-point-prediction.git
cd melting-point-prediction

# Run the notebook
jupyter notebook ensemble-ml-for-melting-point-prediction.ipynb

📁 Project Structure

melting-point-prediction/
├── README.md
├── LICENSE
├── .gitignore
├── ensemble-ml-for-melting-point-prediction.ipynb  # Main solution notebook
└── submission.csv  # Final predictions

🔧 Key Dependencies

• rdkit - Molecular descriptor calculation
• pandas, numpy - Data manipulation
• scikit-learn - ML utilities
• lightgbm - LightGBM model
• xgboost - XGBoost model
• catboost - CatBoost model
• optuna - Hyperparameter optimization
• scipy - Weight optimization
• matplotlib, seaborn - Visualization

📝 Key Techniques

1. SMILES Canonicalization: Standardizing molecular representations
2. Comprehensive Feature Engineering: 1200+ molecular features
3. Ensemble Learning: Combining multiple gradient boosting models
4. Optimal Weight Finding: Using scipy optimization for ensemble weights
5. 5-Fold Cross Validation: Robust model evaluation

🏆 Competition Link

Kaggle: Thermophysical Property - Melting Point

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

👤 Author

Aditya Pawar

• GitHub: @adityapawar327
• Kaggle: Profile

🙏 Acknowledgments

• Kaggle for hosting the competition
• RDKit developers for the excellent cheminformatics library
• The machine learning community for open-source implementations

---

⭐ If you found this helpful, please star the repository!