Surrogate Model Development for Digital Experiments in Welding

Zeyuan Miao ¹, Anastasia Vasileiou ¹, Hujun Yin ¹

⁰School of Engineering, University of Manchester.

Journal of Visualized Experiments : Jove +

|

April 14, 2025

View abstract on PubMed

Summary

This study introduces an automated workflow using artificial neural networks to predict welding-induced residual stress, significantly reducing simulation time and improving accuracy for enhanced structural integrity in manufacturing.

Area Of Science

Materials Science and Engineering
Computational Mechanics
Manufacturing Processes

Background

Welding is critical in manufacturing, but welding-induced residual stress impacts structural integrity.
Predicting residual stress is vital for reliable welded structures.
Traditional simulations are time-consuming, hindering rapid assessment.

Purpose Of The Study

To develop an efficient workflow for predicting welding-induced residual stress using artificial neural networks (ANNs).
To automate data generation from finite element simulations for ANN training.
To reduce the time and effort associated with traditional simulation methods.

Main Methods

Constructing and validating a standard weld finite element simulation.
Developing Python scripts with macro functions for automated data generation.
Training and testing an ANN-based surrogate model on generated data.
Validating simulation results against experimental data.

Main Results

The automated workflow significantly reduces simulation setup and data extraction time.
ANN surrogate models achieved high accuracy in predicting residual stress.
The relative root mean square error was 0.0024, indicating close alignment with simulation results.

Conclusions

The developed workflow offers an efficient and reproducible method for predicting welding-induced residual stress.
ANN-based surrogate models provide a fast and accurate alternative to traditional simulations.
This approach enhances the reliability and durability of welded components in manufacturing.