TY - JOUR
T1 - Predicting light-matter interaction in semi-transparent elliptical packed beds using hybrid deep learning (HDL) approach
AU - Tabassum, Farhin
AU - Eghtesad, Amirsaman
AU - Hajimirza, Shima
N1 - Publisher Copyright:
© 2023 The Authors
PY - 2023/9
Y1 - 2023/9
N2 - In this study, we develop and train a domain-expert supervised learning framework for reliable prediction of the radiative properties of heterogeneous porous media. Our chosen model is based on a hybrid deep learning (HDL) approach employing convolutional neural networks (CNN), thereby extracting in-vitro geometric features, and an artificial neural network (ANN) to concatenate in-situ features as well as physical features while training. Our work is based on a framework of ground truth data generation where an in-house, fine-grained pixel-based Monte Carlo ray tracing (MCRT) simulation algorithm is used for data annotation, and an efficient custom packing algorithm is used for polydisperse and heterogeneous random input configuration generation of elliptical particles. Two distinct CNN architectures, namely LeNet and VGG-16 have been used as hybrid deep learning model cores, are trained to estimate macro radiative properties (e.g., reflectivity, transmissivity, absorptivity) and evaluated. The HDL model with a LeNet core achieves remarkable accuracy, predicting absorptivity, reflectivity, and transmissivity with R2 values exceeding 0.99, 0.93 and 0.63 respectively. Meanwhile, the HDL with a VGG16 core can predict absorptivity, reflectivity, and transmissivity with the accuracy of R2>0.94, R2>0.86, and R2>0.62 respectively. Our findings suggest that while the necessary directional geometric features can contribute to higher accuracy in predictions using the HDL model, achieving greater predictability in transmission requires incorporating more information related to particle materials and molecular structure during training. The results demonstrate that the trained model can reliably be applied to estimate radiation in densely packed porous beds with arbitrarily shaped particles, polydispersity, and varied eccentricity.
AB - In this study, we develop and train a domain-expert supervised learning framework for reliable prediction of the radiative properties of heterogeneous porous media. Our chosen model is based on a hybrid deep learning (HDL) approach employing convolutional neural networks (CNN), thereby extracting in-vitro geometric features, and an artificial neural network (ANN) to concatenate in-situ features as well as physical features while training. Our work is based on a framework of ground truth data generation where an in-house, fine-grained pixel-based Monte Carlo ray tracing (MCRT) simulation algorithm is used for data annotation, and an efficient custom packing algorithm is used for polydisperse and heterogeneous random input configuration generation of elliptical particles. Two distinct CNN architectures, namely LeNet and VGG-16 have been used as hybrid deep learning model cores, are trained to estimate macro radiative properties (e.g., reflectivity, transmissivity, absorptivity) and evaluated. The HDL model with a LeNet core achieves remarkable accuracy, predicting absorptivity, reflectivity, and transmissivity with R2 values exceeding 0.99, 0.93 and 0.63 respectively. Meanwhile, the HDL with a VGG16 core can predict absorptivity, reflectivity, and transmissivity with the accuracy of R2>0.94, R2>0.86, and R2>0.62 respectively. Our findings suggest that while the necessary directional geometric features can contribute to higher accuracy in predictions using the HDL model, achieving greater predictability in transmission requires incorporating more information related to particle materials and molecular structure during training. The results demonstrate that the trained model can reliably be applied to estimate radiation in densely packed porous beds with arbitrarily shaped particles, polydispersity, and varied eccentricity.
KW - Deep learning
KW - Directional geometric features
KW - Energy
KW - Materials
KW - Porous media
KW - Radiative transfer
UR - http://www.scopus.com/inward/record.url?scp=85169502436&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85169502436&partnerID=8YFLogxK
U2 - 10.1016/j.rineng.2023.101368
DO - 10.1016/j.rineng.2023.101368
M3 - Article
AN - SCOPUS:85169502436
VL - 19
JO - Results in Engineering
JF - Results in Engineering
M1 - 101368
ER -