A computational framework based on FEA, ML and GA for estimation of welding residual stresses

The present work describes a computational framework based on finite element (FE) analysis and machine learning (ML) and genetic algorithm (GA) to accurately estimate and minimize the residual stresses in welding. The FE analysis comprises of sequentially coupled thermal and mechanical analyses that are nonlinear due to temperature-dependent thermal/mechanical properties. In the FE analysis, the effective stress function (ESF) algorithm is used for accurate computation of the temperature-induced elasto-plastic stress field during welding. The ESF algorithm accounts for the large variation of yield curves due to temperature changes in welding. To improve the accuracy of the thermal and mechanical analyses, an optimization process combining supervised machine learning and binary coded genetic algorithm is utilized. This optimization technique is implemented to derive an accurate phase-change model for the thermal FE analysis of SS304, which is otherwise a burdensome and costly task to obtain from calorimetric experiments. The results obtained through the derived phase-change model are validated through welding time-temperature distribution reported for a welding experiment in literature. The same optimization process is further used to obtain a set of weld operating conditions (such as weld voltage, current, welding speed, and the gap between plates) for SS304 that reduce the tensile residual stresses. It is found that the accurate estimation and reduction of welding residual stresses can be realized by using the present computational framework.