Mapping regime transitions in stenotic flows using physics-guided active learning

Resolving the transition from laminar to turbulent flow, alongside non-Newtonian shear-thinning effects, in stenosed internal geometries requires highly detailed computational fluid dynamics (CFD). However, the huge computational cost of these simulations makes it difficult to conduct large-scale studies. To overcome this limitation, this study introduces a physics-guided active learning framework designed to efficiently model complex transitional flows. The approach begins with using low-cost baseline laminar simulations and progressively includes complex transitional physics captured via the transition SST model and Carreau–Yasuda rheology through an uncertainty-driven sampling loop. By using a random forest ensemble to estimate areas of high predictive uncertainty, the framework intelligently selects only the most informative flow configurations for expensive, high-fidelity CFD simulation. Applied to a parametrized bifurcating carotid geometry, the resulting surrogate representation reproduces key hemodynamic quantities with a global coefficient of determination of R2 = 0.9869 while requiring roughly 80% fewer high-fidelity simulations than conventional sampling strategies. Analysis of the learned response surface reveals a distinct hydrodynamic transition near a stenosis severity of about 72.5%, corresponding to a local Reynolds number exceeding ∼5200. Above this level, the flow undergoes a qualitative change from smooth acceleration to a jet-dominated state accompanied by shear-layer instability and rapid amplification of wall shear stress. This suggests that stenotic flows may be governed by discrete regime boundaries rather than varying smoothly with geometric narrowing. More broadly, the results show that targeted, physics-informed sampling can provide an efficient route for identifying instability structure in complex internal flows.