Rapid thickness distribution prediction of superplastic formed parts based on geometry adapted PSO-BP neural network surrogate model

Rapid and accurate prediction of thickness distribution remains a critical challenge in achieving real-time process optimization for superplastic forming (SPF) operations. Conventional prediction method based on finite element analysis (FEA) faces constraints in computation efficiency and completely dependency on precisely defined boundary conditions, rendering them unsuitable for real-time control systems. This study proposes a mesh-informed neural network surrogate model based on particle swarm optimized-back propagation (PSO-BP) algorithm to predict the thickness distribution of superplastic forming parts with different geometric feature parameters. A geometric fitness function based on SPF part features is proposed to solve the problem of large local errors in round-corner areas. The mean absolute percentage error of the improved algorithm prediction results has reduced from 1.3% to 0.8% by approximately 38.5% compared to the standard PSO-BP neural network. A rapid prediction of the thickness nephogram of Ti-6Al-4 V box-shaped parts within 0.5 s was achieved with an average deviation from the finite element simulation results less than 1%. This computational advancement enables closed-loop process control by bridging the temporal gap between simulation-based optimization and actual manufacturing cycle times. The developed system shows significant potential for in-process quality monitoring and dynamic parameter adjustment in industrial SPF applications.