基于神经网络的混杂SiC颗粒增强铝基复合材料力学性能预测  

Prediction of Mechanical Properties of Hybrid SiC Particle-reinforced Aluminum-based Composites Based on Neural Network

作　　者：李晓童 庄乾铎 牛志亮 王锶杰 邢正 李赞[2] 岳振明[1] LI Xiaotong;ZHUANG Qianduo;NIU Zhiliang;WANG Sijie;XING Zheng;LI Zan;YUE Zhenming(School of Mechanical,Electrical and Information Engineering,Shandong University(Weihai),Shandong Weihai 264209,China;State Key Laboratory of Metal Matrix Composites,Shanghai 200240,China)

机构地区：[1]山东大学(威海)机电与信息工程学院,山东威海264209 [2]金属基复合材料国家重点实验室,上海200240

出　　处：《精密成形工程》2024年第4期95-100,共6页Journal of Netshape Forming Engineering

基　　金：国家自然科学基金(52175337,52192591)。

摘　　要：目的提高混杂SiC颗粒增强铝基复合材料的韧性,利用卷积神经网络预测其力学性能,以得到力学性能关键因素的影响规律。方法首先,通过实验得到了铝基复合材料的力学性能数据。其次,基于相场裂纹扩展本构,采用Python代码批量生成了不同构型参数的代表性体积单元,并利用Abaqus软件进行了有限元仿真(FEM)。通过代码实现了建模与仿真的一体化构建,利用得到的仿真数据,建立了神经网络模型,并实现了对复合材料力学性能的预测。建模前,对数据进行预处理和筛选,以提高数据质量并降低模型复杂度。最后,建立卷积神经网络,并优化模型的超参数。结果通过建立的神经网络模型,实现了对复合材料力学性能的有效预测。极限强度的预测误差保持在−7%~8.5%,能耗的预测误差保持在−5%~6%,预测精度较高。结论通过结合实验、仿真和卷积神经网络模型,可以更有效地预测混杂SiC颗粒增强铝基复合材料的力学性能,从而为材料设计和制备提供指导。The work aims to enhance the toughness of hybrid SiC particle-reinforced aluminum-based composites and predict the mechanical properties of the composites by utilizing a convolutional neural network(CNN)to determine the key factors affecting their mechanical performance.Firstly,experimental data on the mechanical properties of the aluminum-based composites were obtained.Then,based on the phase-field crack propagation constitutive model,representative volume elements(RVEs)with different configuration parameters were generated by Python code,and finite element simulations(FEM)were conducted with Abaqus software. The integrated construction of modeling and simulation code was realized and the neural networkmodel was constructed with the obtained simulation data, enabling the prediction of the mechanical properties of the composites. Prior to modeling, the data were preprocessed and selected to improve data quality and reduce model complexity. Aconvolutional neural network was established, and the hyperparameters of the model were optimized. The developed neuralnetwork model achieved effective prediction of the mechanical properties of the composites. The prediction error for ultimatestrength ranged from −7% to 8.5%, and for energy absorption ranged from −5% to 6%, demonstrating high prediction accuracy. By combining experiments, simulations, and convolutional neural network models, the mechanical properties of hybridSiC particle-reinforced aluminum-based composites can be predicted more effectively, thereby providing guidance for material design and fabrication.

关 键 词：混杂SiC颗粒 铝基复合材料 卷积神经网络 力学性能预测 相场裂纹扩展本构 

分 类 号：TG1[金属学及工艺—金属学]