A physics-informed neural network to model urea clearance in a hemodialyzer

Hemodialyzers are designed to replicate the function of the kidneys and filter a patient’s blood through a semi-permeable membrane. The clearance of unwanted chemical species, such as urea, depends on the membrane properties and the dialysate and blood flow rates. Here, we measured an Optiflux® F180 dialyzer (Fresenius Medical Care) using a Stemi 2000 Zeiss stereomicroscope and high precision calipers and found that it had approximately 12,600 hollow fibers, a dialyzer inner diameter of 49.53 mm, a hollow fiber inner diameter of 185 μm, a membrane thickness of 36 μm, and a dialyzer length of approximately 266 mm. The geometric complexity of the device precludes performing fully three-dimensional multiphysics simulations involving fluid mechanics and transport of species. Following Karniadakis et al. (Nature Reviews Physics, 2021), we used a physics-informed neural network to model the urea clearance in the model hemodialyzer. We enforced the continuity equation, the Navier-stokes equations and the advection-diffusion equation by training the network on additional information obtained with less computationally expensive results, including those obtained by reducing the number of hollow fibers, which were obtained from finite element analysis software. The neural network model successfully reproduced known clinical urea clearance rates and thus may serve as a useful tool for designing hemodialysis membranes and optimizing flow conditions for medical conditions such as acute kidney injury (AKI).