Surrogate Model Development for Digital Experiments in Welding

The manufacturing industry heavily relies on welding processes to join materials, forming integral components across various sectors. Many aspects will influence the quality of the weld and finally affect the structure integrity of the weldment. Welding-induced residual stress, a consequence of the thermal cycling inherent in these processes, significantly impacts the structural integrity and performance of fabricated components. Understanding and predicting this residual stress is crucial for enhancing the reliability and durability of welded structures. However, quickly assessing a welding setup in digital experiments presents significant challenges, as a traditional simulation may be time-consuming. This research outlines the application of a workflow to construct an artificial neural network-based surrogate model to predict welding-induced residual stress. The model is built using data automatically generated from finite element simulations through Python scripts based on macro functions. Unlike traditional methods that rely on manual preprocessing and finite element simulations, this approach significantly reduces the time and effort required for simulation setup and data extraction, enhancing overall efficiency. By ensuring that all simulation steps are performed consistently through macro functions, the method eliminates human-induced variability, leading to improved reproducibility. Furthermore, the automation of data generation makes it possible to create extensive datasets necessary for training machine learning models, overcoming the limitations of labor-intensive traditional techniques. The workflow contains four main steps: build a standard weld finite element simulation and validate the finite element simulation results against experimental data; develop scripts for large dataset generation using a macro function that records the preprocessing and postprocessing steps of the finite element simulation; use the scripts to generate the required data; develop surrogate models and test their performance. The artificial neural network demonstrated high accuracy in predicting stress levels, closely aligning with simulation results on the test dataset and a relative root mean square error of 0.0024.