Micro-CT-informed physics-conditioned multimodal deep learning for drag coefficient and Nusselt number prediction of irregular coal char particles

Accurate gas–solid flow prediction in combustion and gasification requires drag coefficient (Cd) and Nusselt number (Nu) closures valid for irregular, non-spherical particles, which idealized-shape and scalar-descriptor correlations cannot fully represent. This study predicts Cd and Nu directly from 3D micro-CT images of coal char particles using a physics-conditioned multimodal network supervised by particle-resolved CFD. Ten scanned particles were simulated over six attack directions, Re = 10–200 and particle surface temperatures of 900–1600 K, giving 1080 paired labels; a 3D convolutional encoder learns morphology features that the operating variables modulate through feature-wise linear modulation, avoiding handcrafted descriptors. Generalization is assessed with five disjoint particle-wise folds that hold out every particle exactly once. The model attains Cd R 2 = 0.9517 ± 0.0261 and Nu R 2 = 0.9516 ± 0.0442, with 96.4% of Cd and 99.4% of Nu predictions within ±20% relative error, reducing the mean absolute percentage error of the sphere and Ranz–Marshall correlations by factors of 3.9 and 2.9 on the same rows. A descriptor-only control showed comparable Nu performance within the present morphology range, whereas the full 3D representation provided substantially stronger drag prediction, particularly when unseen particles extended beyond the descriptor range represented during training. These results demonstrate the value of image-based volumetric morphology for orientation-sensitive momentum transfer and generalization to previously unseen irregular particle shapes.