This study presents a hybrid framework integrating Active Learning (AL) and Computational Fluid Dynamics (CFD) to develop efficient surrogate models for predicting centrifugal pump performance across Newtonian and non-Newtonian fluids. Addressing the high computational cost of CFD simulations and the importance of working with the most informative data, the methodology begins with a baseline dataset of synthetic fluids. It expands through AL sampling strategies: input diversity, output novelty, combined input-output exploration, and predictive uncertainty. Two Machine Learning (ML) models, Extreme Gradient Boosting (XGBoost) and Gaussian Process Regression (GPR), were trained and iteratively augmented, with performance evaluated against a fixed test set based on real experimental pump data. Among the sampling strategies, the GPR-based variational method yielded the most stable and accurate improvements, outperforming random selection despite using the same number of training samples. This validates the impact of informed sampling in reducing simulation demands while preserving model fidelity. The study also examines key challenges, including CFD convergence, rheological diversity, and normalisation-induced variability. These findings demonstrate that active sampling strategies can significantly enhance predictive accuracy and generalisation in data-scarce simulation-based workflows for pump modelling.
Morantes-Morales et al. (Sat,) studied this question.