Surrogate-Assisted Multi-Objective Optimization of an Axial-Flow Ventricular Assist Device Rotor Using CFD, Hybrid Surrogates, Sobol Analysis, and MOPSO.

Background

This study presents a surrogate-assisted workflow for multi-objective optimization of an axial-flow ventricular assist device (VAD) rotor using 4809 steady-state CFD simulations spanning six geometric variables.

Methods

Repeated learning-curve analysis and blade-count-stratified Gaussian-process regression (GPR) assessed data efficiency and complemented tree-based surrogate models. Final selection retained GPR for pressure rise, torque, hemolysis index, platelet activation, and exposure time, and HistGradientBoosting for hydraulic efficiency and the low-shear fraction S < 50 Pa, with held-out R2 values of 0.968-1.000. Because hydraulic efficiency is derived from pressure rise and torque at the fixed operating condition, the primary MOPSO formulation used six objectives: maximizing pressure rise and S < 50 Pa while minimizing torque, hemolysis index, platelet activation, and exposure time; efficiency was retained as a secondary engineering metric. Ten independent MOPSO runs were pooled to characterize Pareto trade-offs, and five representative Pareto designs were re-evaluated by direct CFD.

Results

The balanced compromise used a 53.58° inlet blade angle, 20.93° outlet blade angle, three blades, a 0.222 mm clearance gap, 0.796 mm blade thickness, and a 20.00 mm rotor length. For the balanced compromise, mean surrogate-to-CFD relative error was 4.0%, with all individual errors below 10%. Relative to baseline, direct CFD confirmed reductions of 17.2% in torque, 55.6% in hemolysis index, 46.9% in platelet activation, and 27.0% in exposure time, together with a 4.5% increase in S < 50 Pa; pressure rise decreased by 22.4% and hydraulic efficiency changed by +0.7%.

Conclusions

The workflow supports design-space screening, trade-off identification, and targeted CFD confirmation of candidate geometries.