| 摘要: | 在本研究中,針對 107 筆聚苯乙烯高分子玻璃轉化溫度(Glass Transition Temperature, Tg)數據,分別建構多元線性回歸模式與極限梯度提升法(Extreme Gradient Boosting, XGBoost)模式,以比較不同模型在 Tg 預測上的表現差異。所採用的分子描述因子涵蓋了分子結構、拓撲與電子特徵等多種面向,能有效反映分子鏈的化學性質及其對 Tg 的影響。在多元線性回歸模式中,雖能清楚解釋各描述因子與 Tg 的線性關係,但由於難以處理非線性與高階交互作用,其在測試組的預測效果仍存在一定的偏差。
相較之下,XGBoost 模式透過樹模型的集成學習,能靈活捕捉非線性關係,並在多次實驗中展現穩定且優異的表現。本研究進一步透過超參數搜尋與最佳化,包括樹深度、學習率、樹數量及正則化參數的調整,建構出最佳化模型。在最佳組合下,Tg 在訓練組達到R²=0.971,測試組達到R²=0.970,並維持合理的均方根誤差,顯示其具有優異的泛化能力,且明顯優於線性回歸模式及未經最佳化的模型配置。
本研究亦結合夏普利值解釋(Shapley Additive Explanations,SHAP)方法進行模型可視化與解釋,分析各分子描述因子對Tg預測的相對重要性與貢獻方向。結果顯示,部分拓撲與電子結構特徵對 Tg 的提升具有顯著影響,而某些結構性描述因子則傾向於降低 Tg 。透過這種解釋性分析,不僅提升模型的透明度,也有助於理解分子結構與玻璃化行為之間的內在機制。
整體而言,本研究結果顯示,與線性回歸相比,極限梯度提升法搭配超參數搜尋與可視化解釋,能提供更準確與可靠的 Tg 預測工具,對於高分子材料設計與性能預測具有重要的應用價值。
In this study, a dataset comprising 107 polystyrene glass transition temperature (Glass Transition Temperature, Tg) samples was employed to construct both multiple linear regression (MLR) and Extreme Gradient Boosting (XGBoost) models, with the aim of comparing their predictive performance for Tg. The molecular descriptors used in this work encompass multiple aspects, including molecular structure, topology, and electronic characteristics, which effectively represent the chemical properties of polymer chains and their influence on the glass transition temperature. Although the multiple linear regression model allows for clear interpretation of the linear relationships between individual descriptors and Tg, its predictive performance on the testing set exhibits noticeable deviations due to its limited capability in handling nonlinear effects and higher-order interactions.
In contrast, the XGBoost model, based on an ensemble of decision trees, is capable of flexibly capturing nonlinear relationships and demonstrates stable and superior performance across multiple experiments.Furthermore, hyperparameter optimization was performed in this study, including the tuning of tree depth, learning rate, number of trees, and regularization parameters, to construct an optimized predictive model. Under the optimal hyperparameter configuration, the model achieved an R² value of 0.971 on the training set and 0.970 on the testing set, while maintaining a reasonable root mean square error, indicating excellent generalization capability. The optimized XGBoost model significantly outperformed both the linear regression model and the non-optimized model configurations.
In addition, SHAP (SHapley Additive exPlanations) analysis was employed to visualize and interpret the model predictions, enabling the evaluation of the relative importance and contribution direction of each molecular descriptor to Tg prediction. The results reveal that certain topological and electronic structural features have a significant positive impact on Tg, whereas some structural descriptors tend to decrease Tg. This interpretability analysis not only enhances the transparency of the predictive model but also facilitates a deeper understanding of the underlying mechanisms linking molecular structure to glass transition behavior.
Overall, the results of this study demonstrate that, compared to linear regression, the Extreme Gradient Boosting approach combined with hyperparameter optimization and interpretability analysis provides a more accurate and reliable tool for Tg prediction, offering substantial potential for applications in polymer material design and performance prediction. |