Machine learning based on computational fluid dynamics enables geometric design optimisation of the NeoVAD blades

Area of Science:

• Biomedical Engineering
• Cardiovascular Devices
• Computational Fluid Dynamics

Background:

• Pediatric Left Ventricular Assist Devices (LVADs) are crucial for infants with heart failure.
• The hydrodynamic performance and hemocompatibility of LVADs depend heavily on impeller and diffuser blade design.

Purpose of the Study:

• To optimize the blade design of the NeoVAD, a pediatric axial-flow LVAD, for enhanced pump efficiency.
• To utilize computational fluid dynamics (CFD), machine learning, and global optimization techniques for blade design.

Main Methods:

• CFD simulations of 32 base geometries with 6 million hexahedral elements and a Shear Stress Transport turbulence model.
• Validation of CFD models against experimental pressure-flow and efficiency-flow curves.
• Development of surrogate models (multi-linear regression, Gaussian Process Regression, Bayesian Regularised Artificial Neural Network) for efficient optimization using a Genetic Algorithm.

Main Results:

• The optimized NeoVAD blade design achieved a 5.51% increase in efficiency at the design point.
• This represents a 20.9% performance increase compared to the best-performing base design.
• The developed optimization method proved effective for a single objective function.

Conclusions:

• A novel optimization method combining CFD, machine learning, and global optimization successfully enhanced LVAD blade design.
• The optimized NeoVAD design shows significant potential for improved efficiency in pediatric cardiac support.
• Future work will focus on multi-objective optimization for further advancements.