Toward Robust and Stable Model Selection in Machine Learning for Materials Science via Systematic Hyperparameter Optimization

In materials science, Machine learning (ML) has emerged as a powerful and efficient method for predicting and optimizing material properties. However, the widespread use of default hyperparameter approaches presents challenges in ensuring model reliability. This study systematically analyses the hyperparameter optimization (HPO) framework, using carbon dots (CDs) as a representative material system. Common HPO strategies, including Grid Search, Random Search, and Bayesian Optimization, are comparatively evaluated using cross‐validation. These methods were evaluated to assess their effects on prediction accuracy and computational efficiency. The results reveal that HPO significantly improves model generalization and stabilization under limited data conditions, especially over tree‐based algorithms that are sensitive to parameter adjustment. Furthermore, initial screening with default configurations has also proven necessary for the purposeful implementation of HPO. This study demonstrates the critical role of structured HPO in ensuring reproducible, reliable ML models for materials and provides generalizable insights broadly applicable to theoretical materials research.