Improving predictive modeling in materials science based on data augmentation approach

Abstract

Predictive modeling in materials science is often limited by the small size of experimental datasets, which hampers model robustness and generalization. This work proposes a measurement‑informed data augmentation framework for small tabular datasets in materials science. The framework combines bootstrapping, interpolation, and controlled noise addition to expand training data while preserving the underlying statistical structure and metrological characteristics of the original measurements. It is evaluated on four heterogeneous experimental datasets (two adsorption problems, one cement density dataset, and one dredged sediment compressive strength dataset) using Linear Regression, Support Vector Regression, and XGBoost. Across all datasets, data augmentation yields consistent improvements in predictive accuracy and stability, with XGBoost benefiting the most; for example, the coefficient of determination R² for the Planning dataset increases from 0.29 to 0.90 after augmentation, and the Dredging dataset (dredged sediment) shows a substantial reduction in RMSE compared with the baseline model. Because the datasets are very small, these gains are interpreted together with measurement‑focused analyses of bias, uncertainty, repeatability, and residual behavior to avoid over‑stating generalization. The results demonstrate that simple, structure‑preserving augmentation, when combined with a metrological evaluation, can significantly enhance predictive modeling under severe data scarcity in materials science.