Predicting band gap from chemical composition: a comparative study of machine learning and deep learning models

Predicting the band gap of inorganic semiconductors is crucial for designing materials used in electronic and optoelectronic applications. This study employs a range of machine learning (ML) and deep learning (DL) algorithms i.e. Linear Regression, Random Forest, XGBoost, Artificial Neural Network (ANN), and Graph Neural Network (GNN) to predict band gaps using the expt_gap dataset from the Matminer library, which contains experimental band gap values for 6,354 inorganic semiconductors. Following rigorous preprocessing and extraction of composition-based descriptors via the Magpie featurizer, the models were evaluated using a unified fivefold cross-validation protocol applied consistently to all five model classes. The Random Forest model achieved the highest predictive accuracy with a Mean Absolute Error (MAE) of 0.2836 eV, closely followed by the ANN (0.3321 eV) and XGBoost (0.3575 eV). These results significantly outperform the Linear Regression baseline (MAE: 0.5718 eV) and demonstrate the effectiveness of non-linear ensemble and deep learning architectures in capturing the complex, multi-dimensional relationships between chemical composition and electronic properties. The competitive performance of the GNN (MAE: 0.3978 eV) further highlights the potential of graph-based representations in materials informatics. These findings underscore the capacity of advanced ML and DL frameworks to accelerate materials discovery by providing accurate, scalable, and computationally efficient band gap predictions.