導向式訊息傳遞神經網路於聚合物溶劑交互作用參數之預測

Abstract

聚合物－溶劑交互作用的準確預測對於聚合物加工、藥物傳輸與膜分離等應用具有重要意義。Flory-Huggins 作用力參數（χ參數）為衡量聚合物與溶劑相容性的重要指標，然而，其實驗測定往往成本高昂且耗時。本研究提出一套以導向式訊息傳遞神經網路（Directed Message Passing Neural Networks, D-MPNN）為基礎的機器學習框架，能夠從聚合物之單體結構、溶劑結構、溫度與體積分率等資訊直接預測χ參數。我們系統性地評估不同的分子特徵表示方式，以及經驗公式的整合方式，以優化預測準確性。在多種模型中，納入溫度與體積分率資訊的 D-MPNN-TC-sum 模型展現最佳表現（MAE = 0.092，RMSE = 0.162，R² = 0.926），並優於依賴預先計算之指紋與人工特徵的傳統描述子模型。此外，將 Flory-Huggins 理論融入分類架構後，亦能有效預測聚合物與溶劑的相容性， F1分數達 0.915。進一步透過t-分佈隨機鄰近嵌入法（t-distributed Stochastic Neighbor Embedding, t-SNE）視覺化分析顯示，導向式訊息傳遞神經網路能夠捕捉芳香性與環狀結構等關鍵分子特徵，進一步強化模型對聚合物－溶劑交互作用的理解。本研究突顯導向式訊息傳遞神經網路在分子性質預測上的優勢，並說明體積分率為影響相容性預測的重要因素之一。此方法具有良好的可擴展性與可解釋性，為聚合物科學中的機器學習應用提供一個具體可行的架構，有助於推動數據驅動的溶劑選擇與聚合物設計。本篇研究同步發表在ChemRxiv上。

Accurate prediction of polymer–solvent interactions is important in applications such as polymer manufacturing, drug delivery, and membrane separations. The Flory–Huggins interaction parameter (χ) is widely used as a key indicator of polymer–solvent compatibility; however, its experimental acquisition is time-consuming and costly. In this thesis, we propose a machine learning framework based on Directed Message Passing Neural Networks (D-MPNNs) to predict χ parameters directly from monomer structures, solvent structures, temperature, and volume fraction. The methodology systematically explores molecular representation techniques, pooling strategies, and the integration of empirical equations to optimize prediction accuracy. Among the models, the D-MPNN-TC-sum architecture, which includes both temperature and volume fraction as input features, shows strong predictive performance (MAE = 0.092, RMSE = 0.162, R² = 0.926) compared to traditional fingerprint-based models that rely on precomputed descriptors. Furthermore, by incorporating the Flory–Huggins equation into a classification framework, the model achieves highly accurate miscibility classification, with an F1 score of 0.915. Visualization of the learned representations using t-distributed Stochastic Neighbor Embedding (t-SNE) reveals that the model effectively captures structural features, such as aromaticity and cyclic structures, that influence polymer–solvent interactions. Overall, the study illustrates the potential of D-MPNNs in predicting thermodynamic properties and provides an interpretable approach for data-driven solvent selection and polymer design. This work is also available as a preprint on ChemRxiv.