Patient-Specific Hemodynamics via Hybrid Physics-Neural Surrogates

Zero-dimensional reduced-order models resolve bulk cardiovascular hemodynamics in fractions of a second, but their accuracy suffers from drastic simplifications. This work learns the 0D parameters (resistance, inductance) from vessel geometry using networks trained on high-fidelity 3D simulations, keeping the cost and interpretability of the 0D model.

Reported results: Learned parameters reduce error by at least 50% across aortic, aortofemoral and pulmonary anatomies; for the more complex pulmonary cases, error drops from 30% to 7%. Runtime stays under 2 seconds on a personal laptop.

Engineering takeaway: A general recipe - keep the cheap, interpretable physics model, and let a network learn only the closure coefficients that the simplification got wrong.