The present work describes the usefulness of artificial neural networks (ANN) for predicting residual stresses in a butt-welded plate joint. The motivation for the present work is driven by the recently emerging usage of machine learning techniques to reduce computational cost of finite element (FE) analysis. A sample specimen with unequal plate thickness, often used in aircraft panels and shipyards, is investigated here, in order to consider the effects of plate thickness variation in the prediction process. Datasets for building ANNs are generated from sequentially coupled thermal and elasto-plastic FE analysis of the welded specimens using varying inputs like length, width, thickness of plates, weld power, and welding speed. The longitudinal and transverse residual stresses in the domain are obtained as outputs for every set of inputs. ANN models comprising of different activation functions and hidden layers are trained and tested using inputs of three-dimensional coordinate points of the specimens along with the inputs–outputs of FE simulations. The ANN models are then used to predict residual stresses using new input data, and their validity is investigated through FE simulations. It was observed that predictions of ANN models agree well with the simulations, thereby establishing the ANNs as powerful alternatives to FE analysis of welding. For the welding process and material investigated in the present work, the ANN scheme can be used to predict the residual stresses at a considerably faster speed than the conventional FE computations for varying weld-process parameters and geometric parameters.