Transfer Learning for Polymer Mechanics: A Fusion Approach to Bridge Molecular Dynamics Simulations and Experiments in SSBR.

The stress-strain curve is a key indicator of the mechanical behavior of polymeric materials and plays a vital role in optimizing the performance of solution-polymerized styrene-butadiene rubber (SSBR). Molecular dynamics (MD) simulations enable the investigation of microscale deformation mechanisms, yet their use of unrealistically high strain rates leads to stress values that diverge significantly from experimental results. To address this discrepancy, we proposed a weighted fusion framework that integrates transfer learning with a hybrid long short-term memory-multilayer perceptron (LSTM-MLP) model and the eXtreme Gradient Boosting (XGBoost) algorithm. A dataset of 100 simulated stress-strain curves was generated from 20 distinct SSBR molecular systems across five strain rates, supplemented with five experimental curves for SSBR (grade 2557TH) under varying tensile rates. The model was pretrained on the simulated data and fine-tuned using the limited experimental data, enabling stress-strain predictions consistent with experiments. Comparative analyses against alternative machine learning baselines confirmed the model's superior accuracy. Additionally, correlation analysis revealed how the four structural units of SSBR-styrene, 1,2-butadiene, cis-1,4-butadiene, and trans-1,4-butadiene-influence mechanical behavior, offering theoretical insights for targeted performance enhancement.