文献类型：期刊文献

英文题名：Physics-informed Kolmogorov-Arnold network for remaining creep life prediction in 7050 aluminum alloy

作者：Yan, Jianjun[1];Zhou, Junwei[1];Zhang, Jianrui[1];Zhao, Peng[1];Zhang, Ziang[1];Wang, Weize[1];Xuan, Fuzhen[1]

机构：[1]East China Univ Sci & Technol, Sch Mech & Power Engn, Shanghai 200237, Peoples R China

年份：2025

卷号：327

外文期刊名：ENGINEERING FRACTURE MECHANICS

收录：;EI(收录号：20253218927318);WOS:【SCI-EXPANDED(收录号:WOS:001545794400005)】；

基金：This work was supported by the National Key Research and Development Program of China (No. 2021YFB3702204) .

语种：英文

外文关键词：Remaining creep life prediction; Physics-informed Kolmogorov-Arnold network; Physics informed neural network; Hybrid physics-data driven approach; 7050 aluminum alloy

摘要：In the field of engineering, alloys are susceptible to creep damage, which can lead to accumulation of plastic deformation in materials, initiation and propagation of cracks, and structural failure. Accurate prediction of the remaining creep life (RCL) of metal components is crucial to ensure their safe and efficient operation. To address this challenge, we present a hybrid PhysicsInformed Kolmogorov-Arnold Network (PIKAN) modeling approach for predicting the RCL. A novel recurrent neural network (RNN) cell is proposed to construct the RCL processing model, which combines physics-informed layers with data-driven layers. The physics-informed layers based on a revised Larson-Miller parameter method and Robinson's linear damage accumulation rule are constructed to predict the baseline of the RCL, while the data-driven layers based on Kolmogorov-Arnold Network (KAN) are used to compensate for the prediction bias of RCL caused by complex creep behavior that is difficult to model in physics. A dataset for the RCL of uniaxial tensile creep specimens of 7050 aluminum alloy is established. A RCL prediction model is built based on PIKAN. The model is employed to predict the RCL of 7050 aluminum alloy specimens and is compared with physics-based, data-driven, and hybrid-driven methods. The results indicate that PIKAN can adaptively adjust to different operating conditions and effectively compensate for the bias in RCL prediction, offering significant advantages in accuracy, robustness, and interpretability of remaining creep life prediction and providing an efficient solution for RCL prediction in engineering applications.