Multi-scale modelling of pathophysiological flows using Physics Informed Neural Networks

This research project aims at modelling the multi-scale nature of pathophysiological flows in human circulation using machine learning methods.

The pathophysiology of cardiovascular and respiratory conditions shows in several cases different length-scales of description, both from a temporal and spatial point of view.

This represents a critical issue for modelling and numerical simulations and hinders the possibility to pursue more complex and patient-specific analyses.

The innovative method proposed in this research project is based on the use of Physics Informed Neural Networks (PINNs). Their efficiency in multi-scale analyses has been recently tested with outstanding results.

Surrogate models based on PINNs shows an improvement in the efficiency in benchmark tests of several orders of magnitude (seconds w.r.t. hours). The project is organised in Working Packages (WPs) divided over three years.

The first WP is devoted to the training of PINNs to reproduce the relevant physiological phenomena of various circulation pathology such as fibrosis, calcifications, vascular remodelling, thrombus formation, and dissection.

The second WP will focus on the multi-scale analysis of these phenomena in low dimensional models of human vessels (Starling Resistor). In particular, a 2-way Fluid-Structure Interaction analysis will be performed.

In the last WP the previous results will be extended to patient-specific geometries obtained via Magnetic Resonance Imaging data.