Tailoring gas tungsten arc weld geometry using a genetic algorithm and a neural network trained with convective heat flow calculations

Weld attributes like geometry and cooling rate are strong functions of the welding process parameters such as arc current, voltage and welding peed. A specific weld pool geometry can be produced using multiple sets of these welding variables, i.e., different combinations of arc current, oltage and welding speed. At present, there is no systematic methodology that can determine, in a realistic time frame, these multiple paths ased on scientific principles. Here we show that multiple combinations of welding variables necessary to achieve a target gas tungsten arc (GTA) eld geometry can be systematically computed by a real number based genetic algorithm and a neural network that has been trained with the esults of a heat transfer and fluid flow model. The neural network embodies the power of a numerical heat transfer and fluid flow model of GTA elding, since it can predict the fusion zone geometry, peak temperature and cooling rate and its input and output variables are consistent with the quations of conservation of mass, momentum and energy. A genetic algorithm is used to determine a population of solutions by minimizing an bjective function that represents the difference between the calculated and the desired values of weld pool penetration and width. The use of a eural network in place of a heat transfer and fluid flow model significantly expedites the computational task. The desired weld geometry could be btained with various combinations of welding variable sets. The computational methodology described here enables fabrication of a weld with esired geometry within the framework of phenomenological laws via alternative paths involving multiple combinations of welding variables. 2006 Elsevier B.V. All rights reserved.