A systematic study of two machine learning-based approaches for solving non-linear inverse heat conduction problems in one-dimensional domains

In the present study, two machine learning-based approaches using artificial neural networks (ANNs) are presented and compared for solving inverse heat conduction problems (IHCPs) in one-dimensional domains with and without moving boundary conditions. A feed-forward neural network and a nonlinear-autoregressiveexogeneous recurrent neural network (NARX RNN or NARX) are defined based on the digital filter form solution of the IHCPs to facilitate near real-time estimation of surface heat flux using internal temperature measurements. Four cases are considered for the study covering both constant and moving boundary conditions as well as with and without temperature-dependent material properties. Two possible temperature sensor locations are considered, and six different heat flux profiles are used for training and testing purposes. Several robustness studies are also performed, and the performance of each network is assessed based on the accuracy of heat flux prediction under different scenarios. The time required for training each network is also documented. Detailed discussions on the implementation and performance comparison of different networks, sensor locations, and types of heat flux profiles are provided. The results show that both types of ANNs when used in digital filter form offer promising characteristics for solving linear and non-linear IHCPs in one-dimensional domains for cases with and without moving boundary.