Modelling of weld-bead geometry and hardness profile in laser welding of plain carbon steel using neural networks and genetic algorithms

An attempt was made to predict weld-bead geometry and its cross-sectional micro-hardness profile produced by laser welding of plain carbon steel (DC05) for a given set of process parameters. Welding was done using ytterbium fibre laser by considering laser power, weld speed and distance of the focal point from the sample surface as the input parameters. Microscopy was used to measure the weld dimensions. Micro-indentation was made to measure the corresponding Vickers’ hardness along the horizontal cross section. Two different models were developed. The first model had mean hardness and weld-bead geometry represented by four geometrical dimensions of the weld (that is, top width, depth, mid-width and heat-affected-zone width at mid-depth) as the modelling outputs. The second model had the hardness profile plot interpolation parameters as the modelling outputs. Two different designs of neural networks were used for process-based modelling, namely counter-propagation neural network (CPNN) and feed-forward back-propagation neural network (BPNN), and their prediction capabilities were compared. For the feed-forward neural network, a genetic algorithm was later applied to enhance the prediction accuracy by altering its topology. Back-propagation was implemented using 12 different training algorithms. Mean generalisation error was used to compare the modelling accuracy of the neural networks