A generalized characterization of radiative properties of porous media using engineered features and artificial neural networks

At the core of many engineering applications is the evaluation of the radiative responses of materials and devices, particularly at high temperatures. Researchers have focused their attention for decades on the in-vestigation of radiation heat transport (RHT) in porous media. Monte Carlo ray tracing (MCRT) is a reliable computational alternative to the otherwise expensive or infeasible experimental measurements for RHT in heterogeneous media. Despite the accuracy, the computational cost of MCRT simulations can be burden-some given the convergence requirements, the complexity of materials (e.g., large number of particles in a porous medium), and space of input physical specifications (e.g., angle and/or location of incoming rays, wavelength, etc.). This study is an attempt to replace MCRT simulations with supervised learning algo-rithms. Ground truth labeling data is generated using MCRT for random overlapping and non-overlapping circular packed porous media and solving the classical radiative transfer equation (RTE). A significantly low-cost physical and geometrical-based artificial neural network (ANN) model is developed to forecast the radiative characteristics of arbitrary porous configurations, using engineered geometric features. Us-ing the ANN model, a sensitivity analysis is conducted to assess the relative importance of the designed features. This insight helps in identifying the most crucial features to improve the learning for more com-plex geometries. The precision of predictions is calculated based on different hypotheses made over the training data, i.e., whole-size, wall-wise, and pointwise classes. It is shown that when overlapping circles are used as training, the designed model can predict the radiative properties of out-sample data for the first two classes with the accuracy of R 2 > 0 . 944 and R 2 > 0 . 787 , while more directional engineering fea-tures are required to achieve higher accuracies for the former class. Results show that the current model is highly generalizable and applicable to a variety of porous configurations.(c) 2023 Elsevier Ltd. All rights reserved.