A novel systematic numerical approach on determination of heat source parameters in welding process

• Main Authors: Moslemi, Navid, Gohari, Soheil, Abdi, Behzad, Sudin, Izman, Ghandvar, Hamidreza, Redzuan, Norizah, Hassan, Shukur, Ayob, Amran, Rhee, Sehun
• Format: Article
• Language: English
• Published: Elsevier Editora Ltda 2022
• Subjects: TJ Mechanical engineering and machinery
• Online Access: http://eprints.utm.my/103411/1/IzmanSudin2022_ANovelSystematicNumericalApproach.pdf; http://eprints.utm.my/103411/; http://dx.doi.org/10.1016/j.jmrt.2022.04.039

The double-ellipsoidal heat source concept, established by Goldak, has been extensively employed to represent the energy distribution in a broad range of arc welding simulation processes. However, the Goldak's parameters need to be exactly and efficiently defined for accurate arc welding simulation. In this study, a novel procedure was proposed to accurately predict the Goldak's parameters in Gas Tungsten Arc (GTA) Welding simulation. A developed three dimensional (3D) Finite element (FE) analysis was performed to generate thirty sets of normalized input (welding pool characteristics) and outputs (Goldak's parameters). The relevance between Goldak's parameters and welding pool characteristics were established using two regression models and Artificial Neural Network (ANN) computing systems. Linear and quadratic regression models and ANN were compared for evaluation of accuracy of the parameters. Analysis of the results indicated that ANN slightly transcends both regression models, even though the regression models and ANN were able to suitably predict Goldak's parameters for welding simulation. Hence, Goldak's parameters for the welding numerical model were estimated and employed from the ANN model. 3D FE analysis based on thermal-elastic-plastic model using predicted Goldak's parameters were then conducted and validated by the experimental tests in terms of the size of welding pool, temperature distribution and induced residual stress. In the proposed procedure, the data set in training process was obtained using an efficient FE model analysis, which eliminates the cost and time associated with plenty of experimental tests.