In an individualized shee metal assembly line, form and dimensional variation of the in-going parts and different disturbances from the assembly process result in the final geometrical deviations. Securing the final geometrical requirements in the sheet metal assemblies is of importance for achieving aesthetic and functional quality. Spot welding sequence is one of the influential contributors to the final geometrical deviation. Evaluating spot welding sequences to retrieve lower geometrical deviations is computationally expensive. In a geometry assurance digital twin, where assembly parameters are set to reach an optimal geometrical outcome, a limited time is available for performing this computation. Building a surrogate model based on the physical experiment data for each assembly is time-consuming. Performing heuristic search algorithms, together with the FEM simulation, requires extensive evaluations times. In this paper, a neural network approach is introduced for building surrogate models of the individual assemblies. The surrogate model builds the relationship between the spot welding sequence and geometrical deviation. The approach results in a drastic reduction in evaluation time, up to 90%, compared to the genetic algorithm, while reaching a geometrical deviation with marginal error from the global optimum after welding in a sequence.