Semester of Graduation
Fall 2025
Degree
Master of Science in Mechanical Engineering (MSME)
Department
Mechanical and Industrial Engineering
Document Type
Thesis
Abstract
This study addresses the challenge of accurately modeling flows of complex, multiphase media through fluid machines, where highly viscous, compressible, and shear-thinning behavior of the medium complicates the construction of predictive models. Such media include slurries, biomass, foods, and other materials of high value and societal importance. A high-fidelity CFD model of a screw-type fluid machine with an outlet flow restriction is developed. The flowing medium is treated as a single-phase, compressible, non-Newtonian fluid. The model incorporates a non-linear pressure-dependent density model and a strain-rate-dependent viscosity law (Cross power law). A parametric study is conducted by varying key rheological parameters (zero- and infinite-shear viscosity, shear-thinning exponent, time constant) to evaluate their influence on pressure fields and throughput. To efficiently calibrate the rheological model, a Gaussian Process (GP) surrogate is trained on a limited set of CFD simulations and Bayesian optimization techniques are applied. This surrogate-based approach builds a probabilistic model of the expensive CFD objective and uses an acquisition process function to select new parameter combinations, dramatically reducing the number of required simulations for finding minima of the scalar objective.
The results indicate that rheology strongly affects flow. Simulations show that pressure builds near the outlet restriction, while regions of high strain rate concentrate where rotating machine component surfaces are in close proximity to stationary surfaces. Increasing shaft speed reduces pressure due to shear-thinning and increased flow velocity, with mass flow rate linearly increasing with rotational speed. The parametric study indicates that the flow field quantities are highly sensitive to the shear thinning exponent in the Cross model. The surrogate-guided optimization converges quickly to parameter values that produce realistic flow behavior when the true behavior can be measured experimentally. Overall, the machine-learning-enhanced framework identified rheological parameters that produced a closer fit to the experimental measurements, while requiring only a fraction of the computational effort. These findings demonstrate that hybrid CFD–ML methods can effectively capture complex rheology and improve design of feeding systems, paving the way for more efficient renewable-energy feedstock handling.
Date
11-3-2025
Recommended Citation
Saha, Sourav, "Optimization of Rheological Models Using Machine Learning-Based Surrogates" (2025). LSU Master's Theses. 6223.
https://repository.lsu.edu/gradschool_theses/6223
Committee Chair
Owoyele, Ope