From experimental design to data-driven prediction: Modeling the stiffness tunability of freeze-thaw PVA-based...

Mechanically tunable biomaterials play a pivotal role in advanced in vitro modeling platforms because they can be modulated to mimic targeted extracellular matrix (ECM) viscoelastic properties. Specifically, biocompatible hydrogels offer an adaptable and tissue-like platform owning to their interconnected polymer network, high water content, and tunable mechanical properties. However, reliable control of hydrogel tunability remains limited by complex polymer chemistry, water retention, and crosslinking mechanisms. Freeze-thaw PVA-based hydrogels exhibit mechanical properties governed by multivariable and interacting parameters that make their prediction difficult under typical small-data constraints. To address this, we established a modeling framework that couples statistically structured Design of Experiments (DoE) methods with machine learning techniques to capture the linkages between processing parameters and material behavior for stiffness prediction in these polymeric networks. Statistical designs were employed to guide data collection and generate a highly informative dataset for evaluating the effects of polymer concentration, freezing temperature, thawing temperature, and cycle number on hydrogel stiffness. Full factorial and Box-Behnken analyses provided statistically robust findings to assess main effects, interactions, and preliminary regression predictive models. Model-derived findings were screened and experimentally confirmed using differential scanning calorimetry (DSC) and scanning electron microscopy (SEM). After that, we trained, tested, and cross-validated four machine learning models using DoE-based data, achieving improved predictive accuracy compared to DoE regression models. Lastly, we examined a set of synthetic data generated from response surface model to enhance the predictive framework and to improve model robustness and interpolation stability. The integrated DoE-ML framework produced high quality dataset and achieved better predictive accuracy for hydrogel stiffness, with ensemble models outperforming traditional regression, particularly at higher stiffness values where nonlinear effects dominate. Among the evaluated models, Gaussian Process Regression (GPR) approach demonstrated consistently higher R² values across the experimental and synthetic datasets, and improved predictive performance. Overall, this work supports the development of reproducible, data-efficient, and hydrogel-based in vitro models with minimal experimental burden.