Sphere-filling geometric reconstruction for machine learning prediction of drag force on irregular particles

Accurate prediction of aerodynamic drag on irregular particles is essential for modelling gas–solid transport and particle-handling processes. Conventional drag correlations rely on simplified shape descriptors and often perform poorly for angular particles with orientation-dependent flow behaviour. This study develops a geometry-aware machine-learning framework using sphere-filling reconstruction. Three-dimensional particle geometries were obtained by 3D scanning and parameterised using a maximum-inscribed-sphere method. A computational fluid dynamics (CFD) database of 2,000 cases was established for airflow velocities of 20–100 m/s. Wind-tunnel force measurements and particle image velocimetry experiments validated the numerical model, showing an average CFD–experiment deviation of approximately 9.8 %. Five machine-learning models—support vector regression, random forest, a backpropagation neural network, Gaussian process regression, and least-squares boosting—were evaluated. Accuracy improved as the number of filling spheres increased from 10 to 30, confirming the importance of geometric resolution. Least-squares boosting performed best with the 30-sphere representation, achieving R 2 = 0.92 and MAPE = 14 % on an independent test dataset. In comparison, the Wen–Yu, Ganser, and Hoerner correlations yielded R 2 values of only 0.12–0.56. The proposed framework provides an efficient approach for predicting irregular-particle drag and supporting future CFD–DEM simulations of particle transport and pickup.