A FULLY-DENSE DEEP NEURAL NETWORK METHOD FOR THE INVERSE TRANSIENT HEAT TRANSFER PROBLEM

It is necessary to estimate the temperature field accurately during manufacturing and heat treatment processes to perform an effective manufacturing procedure. Under some situations, the front surface temperature can be determined indirectly by the solution of an inverse heat conduction problem (IHCP) based on the transient temperature or heat flux measurements taken at the back surface. This study investigates the capability of a Deep Neural Network (DNN) approach for predicting the front surface temperature and heat flux from the measured back surface parameters. At the early stage, the back surface temperature and heat flux are determined using a direct Python script code. Then, the inverse solution is applied with the help of the fully dense DNN approach. To prevent overfit and non-generalization issues, the regularization and dropout techniques are embedded into the neural network framework. The results reveal that the DNN approach provides more accurate prediction compared to previous mathematical frameworks such as Conjugate Gradient Method (CGM). Moreover, the model is tested by noisy data (from 1% to 10%) causing instabilities in the recovered front surface conditions. Despite of the niose presence, the model can overcome this difficulty and is able to predict the desired parameters with a good accordance.