Transfer Learning-Enhanced Prediction of Glass Transition Temperature in Bismaleimide-Based Polyimides

The glass transition temperature (Tg) is a pivotal parameter governing the thermal and mechanical properties of bismaleimide-based polyimide (BMI) resins. However, limited experimental data for BMI systems posed significant challenges for predictive modeling. To address this gap, this study introduces a hybrid modeling framework leveraging transfer learning. Specifically, a multilayer perceptron (MLP) deep neural network is pre-trained on a large-scale polymer database and subsequently fine-tuned on a small-sample BMI dataset. Complementing this approach, six interpretable machine learning algorithms—Random Forest, Ridge Regression, K-Nearest Neighbors, Bayesian Regression, Support Vector Regression, and Extreme Gradient Boosting —are employed to construct transparent predictive models. SHapley Additive exPlanations (SHAP) analysis is further utilized to quantify the relative contributions of molecular descriptors to Tg. Results demonstrate that the transfer learning strategy achieves superior predictive accuracy in data-scarce scenarios compared to direct training on BMI dataset. SHAP analysis identified charge distribution inhomogeneity, molecular topology, and molecular surface area properties as the major influences on Tg. This integrated framework not only improves the prediction performance, but also provides feasible insights into molecular structure design, laying a solid foundation for rational engineering of high-performance BMI resins.玻璃化转变温度（glass transition temperature, Tg）是调控双马来酰亚胺基聚酰亚胺（bismaleimide-based polyimide, BMI）树脂热学与力学性能的关键参数。然而，现有BMI体系的实验数据较为匮乏，为预测建模带来了显著挑战。为填补这一研究空白，本研究提出了一种融合迁移学习的混合建模框架。具体而言，先在大规模聚合物数据库上对多层感知机（multilayer perceptron, MLP）深度神经网络进行预训练，随后在小样本BMI数据集上完成微调。作为该方法的补充，本研究还采用了六种可解释机器学习算法——随机森林（Random Forest）、岭回归（Ridge Regression）、K近邻（K-Nearest Neighbors）、贝叶斯回归（Bayesian Regression）、支持向量回归（Support Vector Regression）及极限梯度提升（Extreme Gradient Boosting）——构建可解释的预测模型。进一步通过夏普利可加解释（SHapley Additive exPlanations, SHAP）分析，量化了分子描述符对Tg的相对贡献度。研究结果表明，相较于直接在BMI数据集上训练的模型，迁移学习策略在数据稀缺场景下可实现更优异的预测精度。SHAP分析识别出电荷分布不均匀性、分子拓扑结构及分子表面积特性是影响Tg的主要因素。本集成框架不仅提升了预测性能，还为分子结构设计提供了可行的理论指导，为高性能BMI树脂的理性工程化开发奠定了坚实基础。