Transfer learning-enabled viscosity prediction for HAMA/GelMA hybrid hydrogels

Artificial intelligence is transforming the development and design of complex biomedical materials as for example functionalized hydrogels. However, the high experimental costs associated with developing these materials require innovative strategies to reduce data demands for predictive modeling. This study introduces a novel transfer learning approach, termed as Partial Layer Freezing and Re-initialization (PLFRi), designed specifically for small-sample scenarios to predict the viscosity of hybrid hydrogels. Using a multilayer perceptron architecture, we incorporate higher-order nonlinear and weakly nonlinear modules to enable domain adaptation from a precursor system (Hyaluronic Acid/Gelatin) to a target system (Hyaluronic Acid Methacryloyl/Gelatin Methacryloyl). The PLFRi strategy achieves a 119 % improvement in prediction accuracy under limited training data conditions compared to direct modeling. Further optimization of the target-to-source data ratio reveals a trade-off region (7–10 %) between predictive accuracy and cost-efficiency. Additionally, directional sampling of characteristic shear rates enhances model performance (4.56 % improvement in coefficient of determination) and underscores the potential for expanding spatial dimensions into predictive modeling. This study establishes a novel transfer learning paradigm for intelligent hydrogel design, providing a universal and resource-efficient framework for advancing biomaterial development.