ISSN:1000-8365 CN:61-1134/TG
Your Position :Home->Articles Online->2026 Vol.47 No.10

View Full-text
Predicting the High-temperature Compressive Flow Stress of Nickel-basedPolycrystalline Superalloys by a Neural Network
Author of the article:SHEN Wentao1,2, SHANGGUAN Zanpeng3, LIU Qiang1,2,3, WU Ronghai3,4, LUO Hengjun1,2,GUO Lianggang3, LI
Author's Workplace:1. China National Erzhong Group Deyang Wanhang Die Forging Co., Ltd., Deyang 618000, China; 2. Erzhong Group(Deyang) Aviation Science & Technology Co., Ltd., Deyang 618000, China; 3. State Key Laboratory of SolidificationProcessing, Northwestern Polytechnical University, Xi'an 710072, China; 4. Shenzhen Research Institute of NorthwesternPolytechnical University, Shenzhen 518057, China
Key Words:polycrystalline superalloy; high-temperature compression; flow curve; neural network
Abstract:
To simulate the compressive behavior of nickel-based polycrystalline alloys during hot forging, hot compressionexperiments were conducted at deformation temperatures of 880 and 1 030 ℃ at loading rates of 3 and 18 mm/s,respectively. To predict the complex rheological behavior during hot compression, a multi-branch fusion neural network(MBFNN) model was constructed on the basis of the physical characteristics of the material under different deformationmechanisms (work hardening, dynamic recovery, and dynamic recrystallization). This model decouples the effects of strain,strain rate, and temperature into separate branches and then recombines their nonlinear interactions in a fusion layer forprediction. To evaluate the model's predictive performance, traditional fully connected neural network (FCNN) and gated recurrent unit (GRU) models were also established as controls for comparative analysis with the MBFNN model. Theresults show that the MBFNN model significantly outperforms both the FCNN and the GRU models in terms of predictionaccuracy, curve morphology preservation, and physical consistency. It accurately reproduces multi-stage rheological featuressuch as work hardening peaks, steady-state stages of dynamic recovery, and stress softening induced by dynamicrecrystallization. In contrast, the FCNN and GRU models have relatively insufficient fitting capabilities in the softeningstage and peak region and fail to fully reflect the multi-stage evolution of deformation mechanisms.