TY - JOUR
T1 - Learning thermal radiative properties of porous media from engineered geometric features
AU - Hajimirza, Shima
AU - Sharadga, Hussein
N1 - Publisher Copyright:
© 2021
PY - 2021/11
Y1 - 2021/11
N2 - Monte Carlo ray tracing (MCRT) simulations are the most reliable non-experimental means for predicting radiative properties of randomly packed structures or porous media, particularly in polydisperse or heterogenous environments. These methods can approximate radiative properties of a void region filled with homogeneous or heterogenous solids by averaging across a large pool of independent random ray tracing simulations. However, due to stringent precision requirements, high computational cost of simulation setup and execution including processing, memory and programming requirements and parameter-dependent uncertainty of MC simulations, the required time and computational cost of these methods are overwhelmingly high for obtaining accurate estimations. The goal of this work is to build approximate models for MCRT calculations using learning methods. This work is an extension of a previous work, with the goal of building models that are more generalizable and more realistic: while previous work was limited to circular particles and assumed knowledge of the generating parameters of the porous media, the learning models proposed in this work are based on engineered geometric features constructed from the final configuration. Furthermore, ground truth calculations of the current work take the more accurate approach of generating configurations based on a packing algorithm, rather than using the approximate radiation distribution function identification method (pack-free). In addition, we study both uniform and polydisperse 2D media packed with circles and squares of random orientation. We also consider both opaque and transparent materials and study a larger pool of various refractive indices. We focus on the geometric optics dimensions regime in which particle sizes are large compared to the light wavelength and thus the ray tracing simulations render an accurate estimation of the wave propagation properties. We design learning models based on a large synthetically generated database using in-house packing algorithms and MCRT computations. We also use combinations of cross-validation and data mixture techniques to assure that the models are not overfit to a particular class of configurations. As a result, we demonstrate that the models proposed in this work can estimate the radiative properties of configurations in out-sample data generated based on different generation schemes with high accuracy.
AB - Monte Carlo ray tracing (MCRT) simulations are the most reliable non-experimental means for predicting radiative properties of randomly packed structures or porous media, particularly in polydisperse or heterogenous environments. These methods can approximate radiative properties of a void region filled with homogeneous or heterogenous solids by averaging across a large pool of independent random ray tracing simulations. However, due to stringent precision requirements, high computational cost of simulation setup and execution including processing, memory and programming requirements and parameter-dependent uncertainty of MC simulations, the required time and computational cost of these methods are overwhelmingly high for obtaining accurate estimations. The goal of this work is to build approximate models for MCRT calculations using learning methods. This work is an extension of a previous work, with the goal of building models that are more generalizable and more realistic: while previous work was limited to circular particles and assumed knowledge of the generating parameters of the porous media, the learning models proposed in this work are based on engineered geometric features constructed from the final configuration. Furthermore, ground truth calculations of the current work take the more accurate approach of generating configurations based on a packing algorithm, rather than using the approximate radiation distribution function identification method (pack-free). In addition, we study both uniform and polydisperse 2D media packed with circles and squares of random orientation. We also consider both opaque and transparent materials and study a larger pool of various refractive indices. We focus on the geometric optics dimensions regime in which particle sizes are large compared to the light wavelength and thus the ray tracing simulations render an accurate estimation of the wave propagation properties. We design learning models based on a large synthetically generated database using in-house packing algorithms and MCRT computations. We also use combinations of cross-validation and data mixture techniques to assure that the models are not overfit to a particular class of configurations. As a result, we demonstrate that the models proposed in this work can estimate the radiative properties of configurations in out-sample data generated based on different generation schemes with high accuracy.
KW - Monte Carlo Ray-Tracing
KW - Neural network
KW - Porous media
KW - Thermal radiation
UR - http://www.scopus.com/inward/record.url?scp=85111052804&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85111052804&partnerID=8YFLogxK
U2 - 10.1016/j.ijheatmasstransfer.2021.121668
DO - 10.1016/j.ijheatmasstransfer.2021.121668
M3 - Article
AN - SCOPUS:85111052804
SN - 0017-9310
VL - 179
JO - International Journal of Heat and Mass Transfer
JF - International Journal of Heat and Mass Transfer
M1 - 121668
ER -