Surrogate model and machine learning approaches for thermal field reconstruction from weld pool contour: application to GTA welding

Abstract

Thermal cycles in arc welding are crucial as they determine the metallurgy, residual stresses, and distortions of welded parts. Experimentally measuring the temperature everywhere in the welded parts is not possible. This can be achieved with a thermal simulation but finite element analysis requires long computational times, especially for large parts. This study aimed to predict the thermal field using a data-driven approach using numerical and experimental data. First, thermal modeling is defined and arc heating is described with an equivalent heat source. The numerical design of experiments was conducted by varying the heat source parameters. The weld pool contour is extracted from each simulation for building a numerical dataset. The numerical dataset is used for training a surrogate model. The surrogate model is used for estimating the heat source parameters from the weld pool contour using an optimization technique. Then, a K-nearest neighbors algorithm is used to predict the thermal field from the estimated heat source parameters. A significant reduction in computational time is obtained for predicting the thermal field from experimental weld pool contour. Numerical analysis showed that the predicted thermal field is fairly good in the solid than in the weld pool.