Corrosion fatigue life prediction method of aluminum alloys based on back-propagation neural network optimized by Improved Grey Wolf optimization algorithm

• Published in: Journal of materials science Vol. 59; no. 23; pp. 10309 - 10323
• Main Authors: Ji, GaoFei, Li, ZhiPeng, Hu, LingHui, Huang, HaoDong, Song, XianHai, Wu, Qiong
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.06.2024; Springer; Springer Nature B.V
• Subjects: Algorithms; Aluminum; Aluminum alloys; Aluminum base alloys; Back propagation; Back propagation networks; Characterization and Evaluation of Materials; Chemistry and Materials Science; Classical Mechanics; Computation & Theory; Corrosion; Corrosion and anti-corrosives; Corrosion fatigue; Corrosion tests; Crystallography and Scattering Methods; Fatigue; Fatigue life; Fatigue testing machines; Genetic algorithms; Life prediction; Materials; Materials Science; Mathematical optimization; Methods; Neural networks; Optimization; Particle swarm optimization; Polymer Sciences; prediction; Prediction models; Propagation; Root-mean-square errors; Solid Mechanics; Specialty metals industry; wolves
• ISSN: 0022-2461; 1573-4803
• DOI: 10.1007/s10853-024-09799-8

In order to improve the accuracy of the corrosion fatigue life prediction model for the 7050 aluminum alloy, this study presents a corrosion fatigue life prediction model based on back-propagation (BP) neural network optimized by improved grey wolf optimization (IGWO) algorithm that takes into account different sampling orientations and stress magnitudes. The model determines the mapping relationship between corrosion fatigue life, sample direction, and stress level. It also compares the improved grey wolf optimization-backpropagation (IGWO-BP) prediction error with three other models: BP, particle swarm optimization-back propagation (PSO-BP), and genetic algorithm-backpropagation (GA-BP). The IGWO-BP model's performance measures are as follows: a determination coefficient ( R 2 ) of 0.9740, a mean absolute error (MAE) of 0.0479, a root mean square error (RMSE) of 0.0596, and a mean absolute percentage error (MAPE) of 0.9443%. For both the training and test sets, the predicted results are within an error margin of 1.5 times. When compared with the conventional BP neural network, the R 2 of IGWO-BP is increased by 5.69%, and the RMSE, MAE, and MAPE of IGWO-BP are the smallest, reduced by 26.24, 26.19, and 26.72%, respectively.