Numerical Inspection of Heterogeneity in Materials using 2D Heat-Conduction and Hybrid GA-tuned Neural-Network

Present work investigates the effect on heat conduction due to the intrusion in a homogeneous bulk and proposes models to detect its position from the temperature distribution on the surface. Finite volume-based, automated numerical simulations are performed for obtaining the temperature history along/across the bulk surface having different positions of the intrusion. Two approaches are developed to predict the intrusion-position from temperature data. In approach 1, a multi-layer feed-forward neural network (NN) with back-propagation (BP) algorithm is used, whereas the NN parameters are determined through a thorough sequential parametric study. In approach 2, again a NN with BP algorithm is used, but a global evolutionary optimizer, namely genetic algorithm (GA) is employed to optimize the NN parameters. NN with BP algorithm and GA are indigenously developed using 'C' programming language in 'linux' operating system. NN and GA are indigenously combined in a common monolithic platform using some specially designed system commands so that data transfer take place seamlessly in a fully automated way. The performances of the developed approaches are tested and validated in several ways. After comparison, approach 2 is found to have higher prediction capability.