Band Gap Prediction Using Machine Learning Models

Overview

This Colab notebook implements a comprehensive machine learning pipeline to predict the band gap of materials using features extracted from their compositions and structures via the Matminer library. The goal is to accelerate the discovery of new materials by accurately estimating electronic properties with various regression models.

The dataset includes data from the Materials Project API, with features engineered using Matminer featurizers. Several state-of-the-art models were trained and evaluated to compare their performance. Colab

Methodology

1. Data Acquisition & Preprocessing

   • Materials data fetched via Materials Project API
   • Features extracted using Matminer featurizers for compositions and structures
   • Dataset cleaned and NaN features removed

2. Feature Engineering

   • Use of multiple featurizers to capture chemical and structural information
   • Feature concatenation and normalization

3. Model Training & Evaluation

   • Models trained:
     • XGBoost Regressor
     • Gradient Boosting Regressor
     • Random Forest Regressor
     • Stacked Trees (Ensemble)
     • Multi-layer Perceptron (MLP) Regressor
     • Support Vector Regressor (SVR)
     • Linear Regression
     • Ridge Regression
     • Decision Tree Regressor
     • AdaBoost Regressor
     • K-Nearest Neighbors (KNN)
     • ElasticNet
   • Performance evaluated using:
     • Coefficient of Determination (R²)
     • Root Mean Squared Error (RMSE)

---

Results

Initial Model Performance

Model	R²	RMSE
XGBoost	0.8600	0.5528
Gradient Boosting	0.8588	0.5553
Random Forest	0.8452	0.5814
Stacked Trees	0.8327	0.6045
MLP Regressor	0.7847	0.6857
SVR	0.7677	0.7123
Linear Regression	0.6958	0.8150
Ridge Regression	0.6914	0.8209
Decision Tree	0.6563	0.8663
AdaBoost	0.6050	0.9288
KNN	0.3726	1.1705
ElasticNet	0.1324	1.3764

---

Optimized Model Performance

Model	R²	RMSE
Gradient Boosting	0.8827	0.5178
XGBoost	0.8811	0.5213
Random Forest	0.8620	0.5617
Stacked Trees	0.8570	0.5718
MLP Regressor	0.8446	0.5960
SVR	0.8010	0.6745
Decision Tree	0.7250	0.7930
Linear Regression	0.7237	0.7948
Ridge Regression	0.7235	0.7951
AdaBoost	0.6758	0.8609
KNN	0.3734	1.1969
ElasticNet	0.1679	1.3793

---

Usage

1. Clone or open the Colab notebook linked in this repository.
2. Make sure you have an API key for Materials Project (MP API) and set it up in the notebook.
3. Run all cells sequentially to:
   • Download and preprocess the dataset
   • Extract features using Matminer featurizers
   • Train and evaluate machine learning models
4. Modify hyperparameters or featurizers as needed to improve performance.

---

Dependencies

• Python 3.x
• Matminer
• XGBoost
• scikit-learn
• pandas
• numpy

Install all dependencies using:

pip install matminer xgboost scikit-learn pandas numpy