ML-accelerated prediction of radiative view factors in packings of cubical particles

View factors (VFs) are an efficient way to calculate radiative heat exchange in dense and loose packings of particles. Our present study investigates VFs in packings of cubes by means of Monte Carlo Ray Tracing (MCRT) simulations that are precise, but computationally extremely demanding. Subsequently, we use the obtained MCRT data to train and validate machine learning (ML) models to predict VFs based on a plurality of features meant to quantify shadowing. We then demonstrate that our ML models based on (i) a Multilayer Perceptron (MLP), (ii) a graph neural network (GNN), and (iii) a symbolic regressor (SR) can accurately predict VFs, drastically reducing computational cost compared to the MCRT simulations. MLPs achieve the highest accuracy, especially when using physically informed shadowing features, but require significant computation time to calculate the features. The SR relies on the same features, offers a straightforward implementation and achieves slightly lower accuracy. GNNs, using only particle positions, orientations, and direction vectors as features, offer a computationally efficient alternative with competitive performance and strong generalization. Our results show that reliable VF predictions are possible by ML-based models (i) for view factors larger than 10 − 3 , (ii) using easily accessible features, and (iii) at a speed that is many orders of magnitude faster than MCRT. This enables integration of our ML-based models for VFs into large-scale DEM or CFD-DEM simulations of cubical particles. The performance and feature set used in our ML-based models, especially the GNNs, suggests that these approaches are also applicable to arbitrarily-shaped non-spherical particles.