Predicting 3D and skewed velocity and hematocrit distributions in networks of capillary blood vessels using machine learning models

Capillary blood vessels, together with their upstream and downstream vessels known as arterioles and venules, and vascular junctions, form an intricate network. Blood in such small vessels flows as a dense suspension made of red blood cells (RBC) which are extremely deformable. The distribution of blood velocity and hematocrit (RBC concentration) over vessel cross-section is generally complex. The velocity is non-parabolic and often highly skewed. The hematocrit distribution is also radially non-uniform and skewed. A detailed and accurate knowledge of the cross-sectional velocity and hematocrit distribution is of immense physiological importance as these can provide accurate calculation of the wall shear stress and the near-wall cell-free layer. Both in vivo measurements and 1D models of network-scale blood flow have limitations in terms of such detailed information. High-fidelity simulations that treat blood as suspension of deformable RBCs can provide such information. However, such models are computationally very expensive when flow of blood cells in large vascular networks comprised of many vessels and junctions are considered. To overcome this issue, we present machine learning (ML) models that can predict cross-sectional distribution of velocity and hematocrit in large vascular networks. Predictions from our ML models show excellent agreement against the results from high-fidelity RBC-resolved simulations.