Designing and developing high-performance polymeric materials with machine learning has become a major trend in materials science. However, conventional data-driven neural network models often suffer from the "black-box" problem, lacking physical constraints and interpretability, and struggle to achieve accurate predictions with limited experimental data. In this work, a Physics-Embedded Neural Network (PENN) is proposed, which incorporates the Yeoh hyperelastic constitutive model into the network architecture to embed physical laws directly into the learning process, thereby improving physical plausibility and interpretability. The model was first pre-trained on large-scale molecular dynamics (MD) simulation data to capture the intrinsic correlations between polymer structure and mechanical behavior. It was then fine-tuned with a small amount of experimental data through a transfer learning strategy that calibrates the stress magnitude from high-strain-rate simulations to the experimental scale, effectively bridging the gap between simulation and experiment. Uncertainty quantification was further employed to validate the robustness of the predictions. Beyond accurate prediction, PENN enables performance-driven inverse design by mapping target properties to candidate compositional regions, transforming predictive modeling into a practical tool for guiding polymer development.
Zhan et al. (Wed,) studied this question.