Physics-Guided Machine Learning for Robust Viscosity Modeling of HAMA/GelMA Hybrid Hydrogels Under Batch Effect

Reliable viscosity prediction of methacrylated hyaluronic acid (HAMA)/methacrylated gelatin (GelMA) hybrid hydrogels is essential for reproducible multi-batch biofabrication, yet batch effect can severely impair the cross-batch generalization of data-driven models. We established two experimentally measured HAMA/GelMA viscosity datasets collected at different time periods and covering different formulation ranges to characterize batch-effect-induced distributional shifts and support same-batch, cross-batch, and out-of-distribution (OOD) evaluation. Baseline machine learning (ML) models achieved high same-batch predictive performance (R² > 0.90) but degraded markedly under cross-batch conditions (R² = 0.265–0.772). To address these challenges, we propose a physics-informed correction strategy integrating data- and model-level physical priors. A physics-informed data preprocessing (PIDP) strategy filters samples that violate the rheological priors that viscosity increases monotonically with polymer concentration and decreases with temperature. PIDP increased cross-batch R² from 0.713 to 0.892 and reduced RMSE by up to 38.8%. A physics-informed neural network (PINN) further improved physical plausibility of the predictions. Integrating PIDP-PINN strategy (IPPS) achieved the best overall performance (R² = 0.898) and reduced OOD errors by 37.7%. The strategy enables more robust and physically consistent hydrogel viscosity prediction for practical biofabrication, providing a transferable route toward reliable modeling and intelligent design of soft polymer materials under realistic experimental variability.