Physics-informed deep neural network for inverse heat transfer problems in materials

Recently, physics-informed neural networks have shown great promise in providing fast, elegant solutions for inverse problems. In this work, a physics-informed neural network was developed to solve an inverse heat transfer problem through a plane wall of a conductive material where one side is subjected to unknown heat flux, while the other side is thermally insulated. Using our deep neural network, we were able to predict the temperature profiles across the whole material domain and estimate the unknown thermo-physical parameters such as the material’s thermal diffusivity as well as boundary conditions (i.e., heat flux) at the inaccessible side with high accuracy. Furthermore, the method was extended to estimate time-dependent heat flux on the inaccessible side. The predicted temperatures and estimated parameters obtained from our inverse technique are in good agreement with their corresponding exact or true values. For both time-independent and dependent heat flux at the inaccessible side, our model was shown to be very robust in predicting the unknown material properties and/or boundary conditions even from noisy experimental data. The sensitivity of the model’s predictions on the model hyperparameters has also been reported to demonstrate the capability of the proposed method. This method can be utilized for the estimation of multiple thermo-physical parameters using only a few measured data in many practical materials processing involving heat transfer.