A Machine Learning-enhanced Experimental Study of Particle Settling Dynamics in Complex Fluid Systems

• Study of non-spherical particle settling in non-Newtonian fluids • Novel high-precision dataset of 6000 settling experiments • Validated by traditional correlation and state-of-the-art AI model • Machine learning models outperform traditional correlations (R² ≥0.92) • Framework advances AI models for industrial process optimization Understanding particle behavior in complex fluid environments remains a fundamental challenge for optimizing industrial processes across energy (drilling operations), chemical engineering (fluidized bed/gravity settling), food processing (slurry transport), and environmental engineering (sedimentation control) sectors. This study presents an experimental framework that systematically quantifies terminal settling velocity across both Newtonian and non-Newtonian fluid systems using multi-parameter particle characterization. We introduce a new high-precision dataset comprising more than 6,000 data points. It was generated through advanced water column experimentation and high-speed imaging. This study encompasses diverse particle compositions (stone, aluminum, synthetic marble, steel, rubber) with systematically varied equivalent diameters (0.6–1.3 cm), sphericity values (0.67–1.00), and particle densities (1.1–7.9 g/cm³). Our study employs tailored fluid formulations combining sodium chloride, biopolymers, choline chloride, and glycine to achieve controlled rheological property variations. Rheological characterization confirms Herschel-Bulkley model's superiority for capturing nonlinear behavior across all test fluids. The range of generalized Reynolds number for the data set was 10 1 – 10 5 . A representative semi-mechanistic model was applied to the data, confirming its systematic inadequacy in predicting the settling velocity of non-spherical particles in non-Newtonian fluids. In contrast, state-of-the-art machine learning (ML) models demonstrated substantially superior predictive performance. Among these, the Decision Tree Regressor achieved the highest accuracy (R² ≥ 0.92). This study presents the largest and most comprehensive dataset to date for this class of fluid–particle systems. We couple high-fidelity experimental measurements with advanced ML models to achieve substantially improved predictive accuracy for complex fluid–particle interactions. This unified approach facilitates the accurate prediction of complex settling behavior in non-Newtonian fluids.