Surrogate-assisted multi-objective optimization of an axial-flow ventricular assist device rotor using a large CFD-derived design space

Surrogate-assisted optimization can reduce the cost of exploring ventricular-assist-device (VAD) rotor geometries, but the optimization variables, physical constraints, and uncertainty of the surrogate predictions must be treated consistently. This study analyzes a CFD-derived design space of 4809 axial-flow VAD rotors defined by six geometric variables at 5 L/min and 10,000 rpm. Seven archived responses were modeled using output-specific Random Forest, Extra Trees, Gradient Boosting, and radial-basis-function support-vector regressors. Independent test-set R² values ranged from 0.964 to 0.994. Hydraulic efficiency and the archived exposure-time output were retained as diagnostic responses but excluded from optimization because consistency audits did not support their use as independent physical objectives. A constraint-aware multi-objective particle swarm optimization therefore minimized torque, hemolysis index, and platelet activation state and maximized the volume fraction below 50 Pa, while requiring pressure rise to meet the baseline target. Five runs with independent seeds were combined through their best-hypervolume archives, and five diverse Pareto candidates were re-evaluated by direct CFD. The final CFD-confirmed compromise used four blades, a 52.18° inlet angle, a 25.00° outlet angle, a 0.105 mm clearance, a 0.600 mm blade thickness, and a 38.30 mm rotor length. Relative to the baseline, direct CFD maintained pressure rise (+0.15%), reduced torque by 6.77%, increased the low-shear volume fraction by 4.14%, and reduced the comparative hemolysis and platelet-activation indicators by 36.07% and 26.38%, respectively. The results establish a reproducible fixed-operating-point design-screening workflow, while absolute blood-damage interpretation and operation beyond the prescribed point require further validation.