Optimising model parameters is essential for calibrating a Finite Element Model (FEM) of Resistance Spot Welding (RSW) to ensure accurate predictions of weld behaviour and quality. Deterministic optimisation commonly used in current literature, does not account the inherent uncertainty in industrial production. Bayesian model updating provides a framework to quantify these model parameter uncertainties, but its application is limited by the high computational cost of fully coupled electro-thermomechanical simulations. Surrogate modelling is therefore employed in this study to reduce this cost. The surrogate models must predict both scalar outputs (nugget geometry) and time-series signals (dynamic resistance, electrode force) for multi-objective optimization. Due to the high dimensionality and correlation of the time-series signals, dimensional reduction is applied using Sparse Identification of Nonlinear Dynamics (SINDy) [1], which constructs a latent representation. A surrogate model is constructed that maps the welding and model parameters to these latent variables. The original time-series signals are reconstructed by integrating the SINDy-based dynamical system. The methodology is first evaluated on an analytical model using artificial data to evaluate inference accuracy and computational efficiency. This surrogate-assisted workflow enables rapid exploration of key parameters, including contact film properties, electrical and thermal resistivity, and temperature-dependent material behaviour, while preserving the key features of the process signals.