| dc.description.abstract | Predicting the radiative properties of porous media is highly important in many engineering applications (e.g. additive manufacturing). The radiative properties of packed beds highly depend on the geometry and configuration of the structures and the types of materials. The conventional method of computing these properties is through Monte Carlo ray tracing (MCRT) simulations which yield statistical approximations through random sampling of light beams traversing in the porous medium. In ray tracing, numerous light bundles are simulated traveling in random packed beds which are computationally structured via a discrete element model (DEM) of particle settlement simulation. The geometric complexity of porous medium poses computational challenges in both ray tracing and DEM simulations. As a result, MCRT calculations are extremely time-consuming and difficult to setup/program. In this work, we demonstrate that machine learning (ML) techniques can be used to expedite the process of estimating the radiative properties of porous media. ML methods are used in two ways to this aim: 1) As predicative models to directly estimate the radiative properties as functions of the medium geometry and configuration parameters. Specifically, we use neural networks (NN) to predict the radiative properties of the media using supervised learning where the labeling data is collected using ray tracing.
The out-sample prediction can be carried out without the execution of MCRT simulations. We demonstrate that the trained NN models predict transmittivity of random packed beds with improved efficiency and preserved accuracy. 2) As characterization models to summarize and parameterize the statistical geometric properties of random beds which would lead to generation of surrogate penetration length distribution (PLD) functions. PLD is the distribution of probable extinction-free paths in the void between particles and is essential to MCRT simulations. Fast generation of surrogate PLDs essentially obviates the need for cumbersome DEM calculations thus leading to efficient approximate calculations of the radiative properties. ML techniques such as Gaussian Process (GP) modeling can be used for geometric characterization. Coupling ray tracing with the GP model transforms the randomness of the sphere packing into random light travel trajectories in the MCRT simulation. Without DEM simulations, The MCRT coupled with GP model accurately calculates radiative absorptivity. | en |
