Skip to main content

Adrian Bartko

School
Mathematics
Website
LinkedIn
PhD Project

Title: Physics-Consistent Neural Surrogates for Complex Fluid–Structure Interaction

Keywords: fluid–structure interaction | operator learning | vascular modelling | finite element method

My research develops fast neural-network models, known as surrogates, that predict two things about blood flow in the brain's arteries: wall shear stress (the frictional drag of blood on the artery wall) and the pressure waves that travel along the flexible artery walls. Predicting these accurately requires fluid–structure interaction (FSI) simulations, which capture how the blood and the moving wall affect each other, but these simulations are slow. A surrogate learns from a set of them and then gives answers in a fraction of the time. To keep its predictions physically consistent, I design it around the structure of the numerical methods used to solve the FSI equations.

To generate the training data, I have built an automated pipeline that turns artery centrelines traced from brain scans into simulation-ready FSI models. In its first test, it built 55 models from one person's scan in 8.5 minutes.

Supervisors:
Papers:
Conferences:

Impact:

Abnormal wall shear stress (WSS) is associated with where fatty plaques build up and narrow arteries. That narrowing raises the risk of ischaemic stroke, in which the blood supply to part of the brain is blocked. WSS cannot be measured in patients at the resolution needed, so it is usually computed with simulations. Most of these treat the artery wall as rigid, which can overestimate WSS by up to 30%. FSI simulations are more realistic, but with so many arteries in the brain they are too slow for routine clinical use. Fast surrogates trained on FSI simulations are a step towards realistic WSS estimates that are quick enough for the clinic.

Connections
I am keen to connect with researchers, clinicians and industry professionals working on fluid–structure interaction, scientific machine learning, mesh generation or cardiovascular modelling, preferably over a cup of artisan coffee.

Background

I hold an MSci in Mathematics from the University of Glasgow (First Class Honours), where I specialised in applied mathematics, including continuum mechanics, elasticity and field theories. I wrote two theses: one on ergodic theory for data compression, and one on the immersed boundary method for fluid–structure interaction. In the CDT's integrated master's, I simulated how fractal flocs (loose clusters of fine particles) settle in water, using the Nek5000 solver on high-performance computers, in collaboration with Sellafield. My PhD brings these strands together.

Previous Research Interests
Fluid–structure interaction and immersed boundary methods; continuum mechanics and tensor calculus; and the mathematics behind machine learning, including measure and ergodic theory and functional analysis.

Why I chose the CDT in Fluid Dynamics

My dissertation on fluid–structure interaction convinced me to build a career in fluid dynamics. The CDT combines advanced training with close collaboration with industry. It gave me breadth and depth across fluid dynamics before starting my PhD, and experience of working with industry through my project with Sellafield.