Bayesian calibration of resistance spot welding models using temporal process features and streamlined active learning

Accurate multi-physics finite element modelling of resistance spot welding (RSW) remains challenging due to the uncertainties in contact and material parameters that shape the coupled thermal–electrical–mechanical process. This study introduces a Bayesian calibration framework, in contrast to the deterministic parameter optimisation approaches commonly adopted in existing RSW literature, which neglect the inherent variability of the welding process. The proposed framework performs calibration using multivariate model outputs, combining weld nugget geometry with geometrical and temporal attributes (El Ouafi et al., 2012) extracted from process signals such as dynamic resistance and electrode displacement. These features provide physically interpretable information on process evolution while reducing redundancy and correlation relative to full time-series data, enabling more effective parameter identifiability than nugget-only calibration. The methodology is applied to a 2D axisymmetric finite element model of RSW implemented in Simufact. In addition to the electrical and thermal contact parameters, a modified thermal conductivity formulation is introduced and treated as an unknown parameter to be inferred. Bayesian inference is performed using the Streamlined Bayesian Active Learning Cubature (SBALC) method (Li et al., 2025), with experimental multi-physics data (including nugget diameter and height, as well as process signals) used to infer posterior distributions for key modelling input variables.