基于机器学习和数值模拟的选区激光熔化Al-Mg-Sc-Zr合金成形工艺优化  

Forming process optimization of Al-Mg-Sc-Zr alloy by selective laser melting based on machine learning and numerical simulation

作　　者：张喜龙 周永军[1] 郭鹏程 高滕 肖罡 刘筱 朱必武 ZHANG Xilong;ZHOU Yongjun;GUO Pengcheng;GAO Teng;XIAO Gang;LIU Xiao;ZHU Biwu(College of Mechanical and Intelligent Manufacturing,Central South University of Forestry and Technology,Changsha 410004,China;College of Intelligent Manufacturing and Mechanical Engineering,Hunan Institute of Technology,Hengyang 421002,China;Jiangxi Copper Technology Research Institute Co.,Ltd.,Nanchang 330096,China;College of Marine Equipment and Mechanical Engineering,Jimei University,Xiamen 361021,China)

机构地区：[1]中南林业科技大学机械与智能制造学院,长沙410004 [2]湖南工学院智能制造与机械工程学院,衡阳421002 [3]江西铜业技术研究院有限公司,南昌330096 [4]集美大学海洋装备与机械工程学院,厦门361021

出　　处：《中国有色金属学报》2025年第2期487-502,共16页The Chinese Journal of Nonferrous Metals

基　　金：国家自然科学基金资助项目(52475344,52471055,52471132,52475356);湖南省自然科学基金资助项目(2024JJ5644,2022JJ30019);长沙市自然科学基金资助项目(kq2208420)。

摘　　要：本研究旨在利用机器学习方法优化Al-Mg-Sc-Zr高强铝合金的选区激光熔化(Selective laser melting,SLM)成形工艺。由于实验获取样本的成本高、周期长,本文通过SLM成形数值模拟获得充足的实验样本。以激光功率、扫描速度和扫描间距为输入,以数值模拟的残余应力为输出,构建了BP神经网络模型,并利用遗传算法优化BP神经网络模型的预测精度。结果表明:SLM成形Al-Mg-Sc-Zr高强铝合金的残余应力随激光功率的增加而增大,随扫描速度和扫描间距增大而减小。相较于BP神经网络模型,优化的GA-BP神经网络模型的平均绝对误差、均方误差和平均相对误差分别降低了2.62%、14.34%和0.64%。通过搭建遗传算法与GA-BP神经网络的联合双向预测模型,以残余应力最小为目标,确定了Al-Mg-Sc-Zr高强铝合金SLM的最优工艺参数为:激光功率373.17 W、扫描速度919.53 mm/s、扫描间距102.44μm。在最优条件下成形试样的残余应力为201.1 MPa,孔隙率为1.2%,xoz和xoy截面的平均显微硬度分别为103.2HV和120.1HV。本研究可为Al-Mg-Sc-Zr高强铝合金的SLM工艺优化提供一种高效经济的方法。The objective of this study is to optimize the forming process of Al-Mg-Sc-Zr high-strength aluminum alloy by selective laser melting(SLM)using machine learning method.In view of the high cost and long period of obtaining samples through experiments,sufficient samples are obtained through SLM forming numerical simulation.The results show that the residual stress of SLM forming Al-Mg-Sc-Zr high strength aluminum alloy increases with the increase of laser power,while decreases with the increase of scanning speed and scanning spacing.A BP neural network model is constructed with laser power,scanning speed and scanning spacing as inputs and the residual stress from numerical simulation as output,and then optimized using genetic algorithm(GA)to improve its prediction accuracy.Compared with the BP neural network model,the average absolute error,the mean square error and the average relative error of the optimized GA-BP neural network model are reduced by 2.62%,14.34%and 0.64%,respectively.A combined two-way prediction model of GA and GA-BP neural network was established,and the optimal SLM process parameters of Al-Mg-Sc-Zr high-strength aluminum alloy were determined with the minimum residual stress as the objective:laser power 373.17 W,scanning speed 919.53 mm/s and scanning spacing 102.44μm.Based on this process,the residual stress and porosity of the SLM sample are 201.1 MPa and 1.2%,respectively,and the average microhardness on the xoz and xoy sections are 103.2HV and 120.1HV,respectively.This study can provide an efficient and economical method for the SLM process optimization of Al-Mg-Sc-Zr high strength aluminum alloy.

关 键 词：AL-MG-SC-ZR合金 选区激光熔化 机器学习 数值模拟 神经网络 

分 类 号：TG146.2[一般工业技术—材料科学与工程]