Medical devices that come into contact with blood, such as pacemakers, heart valves, catheters, and vessel connections, carry a risk of causing unintentional blood damage. A central concern is hemolysis, where red blood cells rupture under flow-induced stress and release hemoglobin into the plasma, an effect engineers weigh alongside sublethal cell damage and thrombosis when refining medical device design.

To investigate this, engineers at Veryst Engineering, used COMSOL Multiphysics® software and computational fluid dynamics (CFD) to model blood flow through an FDA benchmark nozzle featuring a contraction, neck, and expansion. The team had previously run detailed finite element simulations of this same nozzle, and reused that simulation data to train two deep neural network surrogate models:

• One to predict blood velocity within the nozzle
• Another to estimate hemolysis, using flow rate, converging length, neck diameter, and diverging length as adjustable design inputs

To sharpen accuracy near device walls, the team added a physics-informed model that enforced a no-slip condition, meaning fluid velocity drops to zero right at the wall surface, matching real flow behavior. The surrogate model matched real flow patterns well when the blood flow was turbulent, but struggled more when the flow was smooth and steady (laminar) or shifting between the two states. Its hemolysis predictions, however, stayed close to both the CFD results and published experimental data, and changing the nozzle’s converging and diverging lengths made little difference to hemolysis itself.

Looking forward, Veryst plans to measure exactly how much error the surrogate model introduces and confirm that error is small enough to not change the actual design decision, an approach that could help AI-based simulation gain wider acceptance in future FDA device reviews.

Click to read the original article..