Functional machine learning modeling of electronic bandgap

We present a systematic study of how functional classification of electronic bandgaps improves subsequent machine learning modelings in a group of more than ten thousand semiconductors and insulators. In this regard, we utilize a homemade Python package, MatFeaLib, for systematic generation of 518 descriptors combining 480 composition-based statistics and 38 structural features, and then apply a hybrid parsimonious feature-selection framework to separate about 20 most important features for our final classification, clustering, and regression tasks. Three application-oriented spectral regions, including infrared, visible, and ultraviolet, are selected for our supervised classification tasks along with seven machine learning classifiers, wherein the Extreme Gradient Boost algorithm achieved the best accuracy of 81%, which increased to 94%, after incorporation of lower-fidelity generalized gradient approximation gaps. A hierarchical approach further improves the accuracy of our functional classification to 85% and 95% in the single- and multi-fidelity schemes, respectively. The SHAP analysis revealed electron number, electronegativity, and average bond length as the most influential descriptors, providing physical insights about the classification process. Class-conditioned bandgap regression exhibits a significant improvement of about 40% relative to a global regression. The relevant error diagnostics indicate that different classes may require distinct modeling approaches to capture the underlying relationships and noise characteristics. Unsupervised clustering with an iterative feature-selection technique is used to identify possible hidden patterns in the dataset, which may improve the subsequent multiclass classification and regression procedures. Our findings may open a new avenue for more accurate materials modeling and, thus, more efficient functional materials discovery.