aaburakhia/polymer-property-prediction

Polymer Property Predictor: A Materials Informatics Project

This project is a comprehensive, end-to-end machine learning pipeline designed to predict key physical properties of polymers from their chemical structure.

🎯 Project Overview

The goal of this project is to accelerate materials discovery by building a robust Quantitative Structure-Property Relationship (QSPR) model. The application takes a polymer's structure as a SMILES string and predicts five key properties:

• —Tg: Glass Transition Temperature
• —FFV: Fractional Free Volume
• —Tc: Thermal Decomposition Temperature
• —Density: Material Density
• —Rg: Radius of Gyration

Key Findings & Scientific Insights

1. 1.High Performance on Competition Data: Using 5-Fold Cross-Validation, the final XGBoost models achieved a strong Weighted MAE (wMAE) of 0.2129 on the NeurIPS 2025 dataset.
2. 2.Discovery of Dataset Bias: Rigorous validation against well-known commodity polymers revealed that the model is a specialist that does not generalize perfectly to simpler, out-of-distribution molecules. This is a critical scientific finding about the limitations of the dataset.

🛠️ Technical Methodology

This tool was built using a multi-stage process involving data integration from 7 external sources, RDKit-based feature engineering, and training five separate XGBoost models. The application itself is built with Streamlit and deployed via Docker.

For full project details, code, and the training notebook, please see the [GitHub Repository](https://github.com/aaburakhia/polymer-property-prediction).