Bead-on-plate welding of zircaloy-4 (a reactive material) plates was conducted using electron beam according to central  composite design of experiments. Its predictive models were developed in the form of knowledge-based systems in both forward and reverse directions using neural networks. Input parameters considered for this welding of reactive metals were accelerating voltage, beam current and weld speed. The responses of the welding process were measured in terms of bead width, depth of penetration and micro-hardness. Forward mapping of the welding process was conducted using regression analysis, back-propagation neural network (BPNN), genetic algorithm-tuned neural network (GANN) and particle swarm optimization algorithmtuned neural network (PSONN). Reverse mapping of this process was also carried out using the BPNN, GANN and PSONN-based approaches. Neural network-based approaches could model this welding process of reactive material in both forward and reverse directions efficiently, which is required for the automation of the same.The performance of the neural  network models was found to be data-dependent. The BPNN could outperform the other two approaches for most of the cases but not all in both the forward and reverse mappings.