運用機器學習探討固化劑分子結構對於環氧樹脂系統機械性質之影響

Abstract

環氧樹脂以其多元優異的性能在各應用領域中備受矚目，為擁有高機械強度、耐熱性和優異加工性能等優點的高分子材料，迄今為止仍在蓬勃發展，然而應用層面上的不同考量，使同時兼具多面向機械性質的固化劑材料設計成為難題。本研究運用機器學習方法，輔以具大量數據的資料庫，以2000~2023年DGEBA的機械性質數據為基準，建立固化劑結構與機械性質的關聯性，並將結構直接導入模型進行訓練。首先，以資料筆數、材料組數最多的楊氏模量為代表，透過極致梯度提升迴歸樹演算法建立預測模型，得到良好的相關係數(R)，在預測能力上具有較佳表現，其中，影響楊氏模量的有利結構片段為芳香環衍生物、雙鍵氧、胺基及飒基，芳香環衍生物中又以具備二芳香環共軛稠環的固化劑表現較佳，而不利片段則為二級碳，此外，還看見了環氧樹脂的最終性能與固化劑中片段的連接位置、數量相關，對於高性能固化劑我們還發現稠環種類、環重複數目和對稱性對楊氏模量的影響，提升程度依序為萘環、異吲哚、異苯並噻吩、蒽環和菲環，所適合的環重複數目也隨剛性大小改變，介在2~3之間，至於不對稱稠環結構若以支化形式存在則可以對楊氏模量有正面影響，並且與其他文獻比較後，本研究訓練之模型具有預測值更能符合實驗的優勢。接著藉由同樣的方法、條件建立其他機械性質的模型，大部分相關係數(R)也具有良好表現，為了讓本研究的成果更具應用性，我們還將8種機械性質分為兩大類，對於第一類追求剛性的「楊氏模量、彎曲模量、儲能模量、抗拉強度、抗彎強度」可以使用芳香環衍生物、雙鍵氧、胺基提升，若考量儲能模量還可以額外引入柔性二級碳片段；而第二類須同時保有剛性、柔性的「耐衝擊強度、斷裂伸長率、斷裂韌性」則可以使用單鍵氧、胺基、三級碳來提升，芳香環衍生物則是輔以柔性片段後便可使用；若要同時兼具兩大類機械性質則需使用芳香環衍生物、胺基、雙鍵氧、三級碳和單鍵氧片段，並留意維持材料應有的柔性和剛性。本研究利用機器學習低成本、高效率的特點輔助目前以實驗為主的材料開發方式，促進學者獲得更豐富的結構-性質趨勢，並提供材料設計準則作為研究參考，實現機器學習預測在前，實驗驗證在後的新材料創新策略。

Epoxy resins continue to attract significant attention for their diverse and excellent properties, yet designing curing agents with multiple properties becomes a challenging objective due to the different considerations. In this study, we aim to figure out the relationship between curing agents and the mechanical properties by extreme gradient boosting regression tree algorithm. Firstly, we select "Young's modulus, Flexural modulus, Storage modulus, Tensile strength, Flexural strength, Impact strength, Elongation at break, and Fracture toughness" as features to analyze literature data from 2000 to 2023 and directly input structures for model training, with Young's modulus as representative due to its abundant data sets. Our work indicates that the aromatic ring derivatives, double-bond oxygen, amine, and sulfone can promote Young's modulus by rigidity, cross-linking density and steric hindrance, while the secondary carbon, which increases flexibility owing to single-bond rotation, will decrease it. In addition, the final performance is related to the position and quantity of the structural fragments, and the predicted values obtained from us are more in correspondence with the experiments as compared to others, furthermore, for high-performance curing agents, we establish the ranking of different fused rings that improve Young's modulus. And then, we also divide the properties into two categories. For the first category, "Young's modulus, Flexural modulus, Storage modulus, Tensile strength, and Flexural strength", the aromatic ring derivatives, double-bond oxygen and amine can enhance them via rigidity, steric hindrance and cross-linking density, while considering Storage modulus, flexible secondary carbon can be introduced. The second category of "Impact strength, Elongation at break, and Fracture toughness" can be improved by using single-bond oxygen, amine, and tertiary carbon with their combination of flexibility and rigidity characteristics, and aromatic ring derivatives can be used when supplemented with flexible segments. If considering both categories, aromatic ring derivatives, amines, double-bond oxygen, tertiary carbon, and single-bond oxygen are essential for balancing flexibility and rigidity. Our research utilizes machine learning to support experiment-based material development and provides criteria for materials design. Highlighting a strategy for innovative materials, with machine learning prediction at the forefront and experimental verification as a foundational support.