A Genetic Algorithm and Gradient-Descent- Based Neural Network with the Predictive Power of a Heat and Fluid Flow Model for Welding

In recent years, numerical heat and fluid flow models have provided significant insight into welding processes and welded materials that could not have been achieved otherwise. However, these calculations are complex and time consuming, and are unsuitable in situations where rapid calculations arc desired. A practical solution to this problem is to develop a neural network that is trained with the data generated by a numerical heat and fluid flow model. Apart from providing high computational speed, the results of this neural network conform to the basic laws of conservation of mass, momentum, and energy. In the present study, six feed-forward neural networks have been developed for the gas tungsten arc (GTA) welding of low-carbon steel. Each network provides one of the six output parameters of GTA welds, i.e., depth, width, and length of the weld pool, peak temperature, cooling time from 800° to 500°C, and maximum liquid velocity. The networks require values of 17 input parameters including the welding variables like current, voltage, welding speed, arc efficiency, arc radius, and power distribution factor, and material properties like thermal conductivity and specific heat. The weights of the neural networks were calculated using two optimization schemes, first using the gradient descent (GD) method with various sets of randomized initial weights, and then applying a hybrid optimization scheme where a genetic algorithm (GA) is used in combination with the GD method. The 5. M1SHRA is now with Department of Metallurgical Engineering and Materials Science. Indian Institute of Technology, Bombay, India. Previously he was with Department of Materials Science and Engineering, The Pennsylvania State University, University Park, Pa, as is T. Dehroy. BY S. MISHRA AND T. DEBROY neural networks produced by the hybrid optimization approach gave better results than all the networks based on only the GD method. Unlike the GD method alone, the hybrid optimization scheme could find the significantly better weights, which is illustrated by the good agreement between all the outputs from the neural networks and the corresponding results from the heat and fluid flow model.