Live simulation
06 · Computational Mechanics & AI Lab · WashU · May – Aug 2025

CFD + Machine Learning —
Hemodynamic Simulations

1000×
Speedup vs. traditional CFD solvers
3D–0D
Coupled cardiovascular model
4
Cardiac valves simulated (MV, AV, TV, PV)
Python / PyTorchPhysics-Informed Neural Networks CFDRunge-Kutta 4 Fortran → PythonNumerical MethodsHemodynamics

Summer research at WashU’s Computational Mechanics & AI Lab under Dr. Fanwei Kong, mentored by Yusuf Ozturk (Graduate). The project focused on replacing a legacy Fortran-based cardiovascular solver with a modern, ML-extensible Python pipeline — and coupling it with a 3D CFD solver to simulate realistic cardiac hemodynamics.

The video above shows the hemodynamic velocity field through the heart ventricle — the kind of simulation that traditionally takes hours per run. The goal was to make this fast enough for real-time parameter studies and eventual patient-specific modeling.

The core architecture couples a 3D fluid dynamics solver (full Navier-Stokes, finite element) with a 0D cardiovascular circulation model — a lumped-parameter electrical analog of the heart and vasculature. The two solvers exchange pressure and flow data at each timestep through direct function calls, no file I/O.

3D-0D Coupling Diagram
Coupled 3D fluid dynamics + 0D cardiovascular circulation model — pressure/flow interface
Hemodynamic Velocity Field
Hemodynamic velocity streamlines through the cardiac chambers — CFD output

Findings were presented at the WashU Summer Engineering Fellowship (WUSEF) symposium. The poster covers the full pipeline: background motivation, coupled model architecture, implementation details, and validated results showing physiologically accurate valve pressure waveforms.

WUSEF 2025 Symposium
WUSEF Symposium Poster
Coupling 3D CFD with 0D Cardiovascular Models in PyTorch · Luis Garcia · WashU Dept. of Mechanical Engineering
Legacy code is documentation in disguise
The Fortran solver had no comments and no documentation. Reading it line by line to understand what each variable represented taught me more about cardiovascular modeling than any paper. Translating it forced a level of understanding that just using a modern library would have skipped entirely.
PINNs are powerful but not magic
The 1000× speedup is real but comes with tradeoffs — PINNs struggle near discontinuities like valve opening/closing events. Understanding where the physics-based constraints help vs. where they fight the optimizer was the hardest part of the project.
Validation is the whole job
A fast solver that gives wrong answers is worse than a slow solver. Every result had to be checked against physiological reference data and conservation laws before it meant anything. I learned to be skeptical of my own outputs by default.
Design-Build-Fly All projects CAD Portfolio