6-12. Designing epoxy resin formulations by high-throughput calculation and machine learning

Jie Tao¹, Hao Wang¹, Kai Jin²

1. School of materials science and technology, nanjing university of aeronautics and astronautics

2. School of mechatronics, nanjing university of aeronautics and astronautics

Abstract: In this study, a new epoxy resin formulation was designed based on a high-throughput molecular dynamics calculation platform and machine learning. Based on the results of glass transition temperature, Young's modulus, tensile strength and elongation, the resin system with DGEBA/TGDDM/DDS/PES molar ratio of 0.2/0.8/1/0.06 was selected as the high performance resin system and verified by experiments. The glass transition temperature can achieve 540K, the Young's modulus is 3.8GPa, the tensile strength is 73MPa and the elongation is more than 4%, which proves the effectiveness of high-throughput molecular dynamics calculation platform and machine learning in resin development.

Epoxy resin has widely used in electronic appliances, aerospace and other fields due to its advantages of high bonding strength, good dimensional stability, good chemical resistance, high mechanical strength, excellent electrical insulation, and strong radiation resistance. Generally, epoxy resin formulations are classified into three types: matrix, curing agent and modifier, the diglycidyl ether of bis phenol A (DGEBA) and tetraglycidyl diaminodiphenylmethane (TGDDM) are used as the based resin, the dicyandiamide (DICY) and 4,4-diaminodiphenyl sulfone (DDS) were selected as curing agent, the polyetherimide (PEI) and polyethersulfone (PES) were choosed as toughening agents, which were a total of six components. In the research of epoxy resin formula, the experimental trial and error method is usually used to find the resin with superior mechanical properties and functionality. The traditional development model is time consuming, expensive, and inefficient in researching and developing six component epoxy resin formulation. To this end, we proposed a new epoxy resin design method based on high-throughput computing platform and machine learning to screen the optimal resin formulation and use experiments to verify the results.

Based on the high-throughput computing platform (shown in Fig. 1) which has the ability to generate 1800 sets of computing cases in a one-time concurrent manner and can complete the process operation and calculation, we calculated 15 sets of cases based on platform and machine learning. First, complete the concurrent calculation through the high-throughput computing platform, then the calculated result is input as raw data into the neural network of machine learning, and the neural network trains the original data. The six components were used as BP's six variables in the network input layer. The glass transition temperature, Young's modulus, tensile strength and elongation were respectively taken as the four output results of BP's network output layer. Then we test the trained neural network and found that the maximum error is less than 13%. More than 1,300 epoxy resin systems were predicted by this trained neural network model. Based on the comprehensive glass transition temperature, Young's modulus, tensile strength and elongation, the resin system with DGEBA/TGDDM/DDS/PES molar ratio of 0.2/0.8/1/0.06 was selected as the high performance resin system, and the composition of this high performance resin system was verified by experiments. The glass transition temperature can achieve 540 K, the Young's modulus is 3.8GPa, the tensile strength is 73MPa, and the elongation is more than 4%.

Keywords: Material development; High-throughput calculation; Machine learning; Epoxy resin design

Fig. 1 High-throughput resin computing platform start page

图1 高通量树脂计算平台开始页面

Fig. 2 The component calculation results - glass transition temperature (Tg)

图2计算结果组分-玻璃化转变温度

基于高通量计算和机器学习设计环氧树脂配方

陶杰¹，王浩¹，靳凯²

1 南京航空航天大学，材料科学与技术学院; 2 南京南京航空航天大学，机电学院

摘要：在这项研究中，基于高通量分子动力学计算平台和机器学习，设计了新环氧树脂配方。综合玻璃化转变温度、杨氏模量、拉伸强度和延伸率，选取DGEBA/TGDDM/DDS/PES摩尔比为0.2/0.8/1/0.06的树脂体系为高性能树脂体系，并用实验验证了这个高性能树脂体系组成。该树脂体系玻璃化转变温度达到540K，杨氏模量达到3.8GPa，拉伸强度为73MPa，延伸率超过4%,证明了高通量分子动力学计算平台和机器学习在树脂研发中的有效性。

环氧树脂具有粘接强度高、尺寸稳定性好、耐化学药品性好、力学强度高、电绝缘性优良，以及耐辐照性较强等优点，广泛应用于电子电器、航空航天等领域。通常环氧树脂配方分为基体，固化剂和改性剂三大类，双酚A的二缩水甘油醚（DGEBA）和四缩水甘油基二氨基二苯基甲烷（TGDDM）被用作基础树脂，固化剂选取双氰胺（DICY）和4,4-二氨基二苯砜（DDS），增韧剂为聚醚酰亚胺（PEI）和聚醚砜（PES），总计六组分。在环氧树脂配方研发中，通常采用的是实验试错法去寻找力学性能较优以及具有功能性的树脂，在六组分环氧树脂配方中，采用传统研发模式，耗时长，成本高，研发效率低。为此我们提出了新型环氧树脂设计方法，基于高通量计算平台以及机器学习，去筛选最优树脂配方，并用实验去验证相应的结果。

基于高通量计算平台（如图1所示），具有一次性并发式生成1800组计算案例的能力且完成流程化操作及计算，基于平台和机器学习我们计算了15组案例，首先通过高通量计算平台，完成并发式计算，此时将计算的结果作为原始数据输入到机器学习的神经网络中，神经网络会对原始数据进行训练。将六种组分分别作为BP神经网络输入层的六个变量，玻璃化转变温度、杨氏模量、拉伸强度和延伸率分别作为BP神经网络输出层的四个输出结果。对训练好的神经网络进行测试，发现最大误差不超过13%，通过这个训练好的神经网络模型预测了超过1300种环氧树脂体系。综合玻璃化转变温度、杨氏模量、拉伸强度和延伸率，选取DGEBA/TGDDM/DDS/PES摩尔比为0.2/0.8/1/0.06的树脂体系为高性能树脂体系，并用实验验证了这个高性能树脂体系组成。该树脂体系玻璃化转变温度达到540K，杨氏模量达到3.8GPa，拉伸强度为73MPa，延伸率超过4%。

关键词：材料研发；高通量计算；机器学习；环氧树脂设计