Modelling Convective Heat Transfer Coefficient During TIG Welding Using an Artificial Neural Network

Abstract

Tungsten Inert Gas (TIG) welding demands precise control of convective heat transfer coefficient (h) to optimise weld pool dynamics, minimise distortions, and enhance mechanical properties in critical applications like aerospace and automotive sectors. This study develops an Artificial Neural Network (ANN) model for real-time prediction of h using welding current, voltage, and speed as inputs, derived from Central Composite Design (CCD) experiments on 10 mm mild steel plates. A feedforward backpropagation ANN with Levenberg-Marquardt training (trainlm), log-sigmoid activation, and random data division achieved rapid convergence (MSE ~10⁻³ by epoch 7), with validation R=0.915, test R=0.965, and overall R=0.710, yielding mean absolute errors <0.1 W/(m²·K) across 20 runs. Regression analysis confirmed 67.8% R² fit, superior to empirical Nusselt correlations (>20% error), enabling adaptive process control for distortion-free welds. ANOVA (F=40.98, p<0.001) validated model significance, positioning ANN as a robust tool for TIG optimisation.