主動學習高分子材料設計

Abstract

高分子是生活中常見的材料，能夠透過多樣化的材料設計方式得到不同性能，以滿足各種需求，然而過往使用實驗高分子設計所需成本過高，因此有許多研究試圖以數據驅動方法解決此設計難題。現今數據驅動方法蓬勃發展，數據驅動用於高分子設計也成為熱門的研究議題，然而資料不足侷限了其發展，固在本研究中導入主動學習技術，解決此設計難題。 本研究共分為三部分，皆以主動學習進行高分子材料性質設計最佳化。包含使用主動學習進行隨機共聚物序列設計，成功減少98%需要標註的資料量，有望解決標註資料不足的問題。第二部分的研究則是進行單體多目標最佳化，透過主動學習驅動實驗設計，僅以102以下的標註資料量，即可在龐大的設計空間中完成高分子單體多目標最佳化。最後一部分研究將先前的方法用於純實驗數據驅動設計，並以Vitrimer之配方以及比例進行研究，透過模型建議，成功在設計空間中找出機械性質最佳之配方。 本研究解決了過往數據驅動高分子設計的難題，並以模擬及實驗資料分別進行驗證，此研究成果所建立的主動學習設計方法，可望用於實際工業應用，加速不同需求之高分子材料的開發。

Polymers are ubiquitous materials in daily life, offering a wide range of properties through versatile material design strategies to meet diverse application requirements. However, traditional experimental approaches to polymer design often lead to high costs, which motivates recent efforts to address this challenge using data-driven methods. With the rapid advancement of data-driven technologies, their application to polymer design has emerged as a prominent research focus. However, the limited availability of labeled data remains a major bottleneck that hinders further progress. To overcome this limitation, we introduce active learning techniques to enhance polymer design processes. This research is structured into three major parts, each leveraging active learning to optimize the design of polymer materials. The first part involves the design of random copolymer sequences using active learning, which successfully reduces the amount of required labeled data by 98%, thus addressing the problem of data scarcity. The second part focuses on the multiobjective optimization of monomer structures, where active learning is employed to guide experimental design. Remarkably, with fewer than labeled data points, the method achieves effective multi-objective optimization across a vast design space. The final part applies the developed methods to a fully experimental dataset, targeting the formulation and composition optimization of vitrimers. Using model-driven suggestions, the optimal formulation with superior mechanical properties was identified within the design space. Overall, this study addresses key challenges in data-driven polymer design by integrating active learning, validated through both simulation and experimental datasets. The proposed active learning framework holds promise for practical industrial applications, offering a pathway to accelerate the development of polymer materials tailored to specific performance requirements.