Rectangular-duct slurry transport exhibits coupled particle stratification and velocity-profile asymmetry that cannot be represented adequately by circular-pipe correlations. This study develops a validated CFD–machine-learning framework to quantify these effects and predict concentration and velocity profiles. A three-dimensional mixture model coupled with kinetic-theory and turbulence closures examined the influences of velocity, solids concentration, hydraulic diameter, particle density, and particle size. The resulting 81-case full-factorial CFD database trained separate shallow artificial neural networks, with predictive assessment extended to Support Vector Regression, Random Forest, and power-law regression. Increasing velocity from 2.0 to 3.5 m/s reduced the dimensionless bottom-to-top concentration ratio from approximately 3.5 to nearly 1.0, whereas higher concentration, density, and particle size intensified settling and lower-region accumulation. The proposed flux indices showed that the upper duct half consistently transported more streamwise velocity than the lower half. Coarse iron ore produced the strongest asymmetry, with AI_flux = 7.31% and AR_flux = 1.088. Pressure loss increased with velocity, concentration, and density, while larger ducts reduced it by more than 50%. For three independent intermediate cases, the velocity ANN achieved mean R²= 0.99937 and RMSE = 0.00203. The framework enables rapid, CFD-grounded assessment of non-circular slurry systems within the investigated domain and supports design optimisation without repeated full-scale CFD simulations.