Chen Qionghai, Liu Zhanjie, Huang Yongdi, Hu Anwen, Huang Wanhui, Zhang Liqun, Cui Lihong, Liu Jun
Key Laboratory of Beijing City on Preparation and Processing of Novel Polymer Materials, Beijing University of Chemical Technology, Beijing 100029, People's Republic of China.
Beijing Engineering Research Center of Advanced Elastomers, Beijing University of Chemical Technology, Beijing 100029, People's Republic of China.
Langmuir. 2023 Dec 5;39(48):17088-17099. doi: 10.1021/acs.langmuir.3c01878. Epub 2023 Nov 20.
Natural rubber (NR) with excellent mechanical properties, mainly attributed to its strain-induced crystallization (SIC), has garnered significant scientific and technological interest. With the aid of molecular dynamics (MD) simulations, we can investigate the impacts of crucial structural elements on SIC on the molecular scale. Nonetheless, the computational complexity and time-consuming nature of this high-precision method constrain its widespread application. The integration of machine learning with MD represents a promising avenue for enhancing the speed of simulations while maintaining accuracy. Herein, we developed a crystallinity algorithm tailored to the SIC properties of natural rubber materials. With the data enhancement algorithm, the high evaluation value of the prediction model ensures the accuracy of the computational simulation results. In contrast to the direct utilization of small sample prediction algorithms, we propose a novel concept grounded in feature engineering. The proposed machine learning (ML) methodology consists of (1) An eXtreme Gradient Boosting (XGB) model to predict the crystallinity of NR; (2) a generative adversarial network (GAN) data augmentation algorithm to optimize the utilization of the limited training data, which is utilized to construct the XGB prediction model; (3) an elaboration of the effects induced by phospholipid and protein percentage (ω), hydrogen bond strength (ε), and non-hydrogen bond strength (ε) of natural rubber materials with crystallinity prediction under dynamic conditions are analyzed by employing weight integration with feature importance analysis. Eventually, we succeeded in concluding that ε has the most significant effect on the strain-induced crystallinity, followed by ω and finally ε.
天然橡胶（NR）具有优异的机械性能，这主要归因于其应变诱导结晶（SIC），因此引起了广泛的科技关注。借助分子动力学（MD）模拟，我们可以在分子尺度上研究关键结构元素对SIC的影响。然而，这种高精度方法的计算复杂性和耗时性限制了其广泛应用。将机器学习与MD相结合是在保持准确性的同时提高模拟速度的一条有前途的途径。在此，我们开发了一种针对天然橡胶材料SIC特性的结晶度算法。通过数据增强算法，预测模型的高评估值确保了计算模拟结果的准确性。与直接使用小样本预测算法不同，我们提出了一种基于特征工程的新颖概念。所提出的机器学习（ML）方法包括：（1）一个极端梯度提升（XGB）模型来预测NR的结晶度；（2）一种生成对抗网络（GAN）数据增强算法，以优化对有限训练数据的利用，该算法用于构建XGB预测模型；（3）通过采用权重积分与特征重要性分析，阐述了天然橡胶材料中磷脂和蛋白质百分比（ω）、氢键强度（ε）和非氢键强度（ε）对动态条件下结晶度预测的影响。最终，我们成功得出结论，ε对应变诱导结晶度的影响最为显著，其次是ω，最后是ε。
