CFD-AI methodology for translating Brookfield measurements into rheometer data

Sammanfattning

Brookfield viscometers are widely used in industrial quality control because they are simple and practical, but the viscosity values obtained are relative to the instrument, spindle, container geometry, and measurement procedure. In contrast, rotational rheometers such as the Anton Paar RheolabQC provide more defined rheological data under controlled shear conditions. This thesis developed and evaluated a novel workflow combining computational fluid dynamics (CFD) and artificial intelligence (AI) to link Brookfield measurements to rheological model parameters that can be used in CFD simulations. The method combines experimental measurements with the Brookfield DV2T-RV with RV5 spindle and Anton Paar RheolabQC with CC27 geometry, Cross model regression, CFD validation, CFD-generated dataset creation, and artificial neural network prediction into one integrated framework. An initial Power Law-based approach demonstrated that AI could learn the relationship between CFD-generated torque in the RV5 geometry and rheological parameters, but laboratory comparison showed that the Power Law model was insufficient for transfer between the RV5 and CC27 geometries. Shear rate profile analysis showed that the CC27 geometry mainly samples a higher and narrower shear rate range, while the RV5 includes very low shear rates even at high RPM. The Cross model was therefore selected because it can represent the low-shear plateau, transition region, and shear-thinning behaviour. CMC30000 solutions at 1.0, 1.4, and 1.5 wt% were used to span a broad torque range measured using the Brookfield viscometer. Due to higher variability in the 1.5 wt% measurements, the main validation and AI evaluation focused on the 1.0 and 1.4 wt% solutions. CFD validation showed that Cross model parameters regressed from CC27 geometry data reproduced CC27 torque well when end-effect correction was included, reducing the mean torque deviation to approximately 1%. Artificial neural networks were then trained on 5400 CFD-generated RV5 simulations to predict Cross model parameters from Brookfield-type inputs. Both 5-point and 10-point models achieved high accuracy on synthetic test data, but the 10-point model gave lower maximum errors and better torque reconstruction. When AI-predicted parameters were reintroduced into CFD, the 10-point model reproduced measured torque from the Brookfield viscometer with mean deviations of approximately −1.74% to +0.49% for 1.0 wt% and −1.41% to +0.16% for 1.4 wt%. The results show that the developed CFD-AI workflow is a promising proof-of-concept for obtaining CFD-usable Cross parameter estimates from controlled Brookfield measurements. The method is especially valuable because it demonstrates a practical route from simple industrial viscosity measurements to simulation-ready parameter estimates. Although the AI-predicted parameters did not fully match the CC27 geometry–regressed rheological parameters, the torque agreement achieved with the 10-point model indicates strong potential for further development. With improved low-shear measurements, broader CFD datasets, refined uncertainty modelling, and validation on additional fluids and spindle geometries, the workflow could become a useful tool for industrial rheological characterization and process simulation.