The latest articles on DEV Community by Arno Meijer (@arno_meijer). https://dev.to/arno_meijer https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3931034%2Fccacedc0-39a9-4849-a559-025bb84a8dc5.png https://dev.to/arno_meijer en Arno Meijer Thu, 14 May 2026 11:19:12 +0000 https://dev.to/arno_meijer/building-a-real-time-cfd-framework-in-jax-aerojax-1aj7 https://dev.to/arno_meijer/building-a-real-time-cfd-framework-in-jax-aerojax-1aj7 <p>I’ve been building a real-time 2D CFD framework in JAX called AeroJAX.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F47736wotz7f7oml3d9ur.gif" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F47736wotz7f7oml3d9ur.gif" alt="Real-time AeroJAX CFD simulation of an airfoil flow field, visualising wake formation and vortex detection in a JAX-based solver" width="" height=""></a></p> <p>The goal is to make fluid simulation more interactive, so you can explore flow behaviour in real time instead of waiting for batch simulation runs.</p> <p>Most CFD tools are batch oriented. You set up a case, run it, and analyse results afterwards. If you change geometry or parameters, you usually restart the simulation.</p> <p>AeroJAX is built around a different idea: <strong>you modify the simulation while it is running.</strong></p> <p>👉 GitHub Repository: <a href="https://github.com/arriemeijer-creator/AeroJAX" rel="noopener noreferrer">https://github.com/arriemeijer-creator/AeroJAX</a></p> <h3> What it does </h3> <ul> <li>GPU-accelerated CFD using JAX and XLA</li> <li>Image-based SDF geometry input (white = fluid, non-white = solid)</li> <li>Real-time flow field updates during execution</li> <li>Wake and separation visualisation as the simulation evolves</li> <li>Swappable solver components (pressure solvers, LES models, etc.)</li> <li>Optional neural operator experiments inside the loop with in-program training</li> </ul> <p>You can draw or move obstacles during the simulation and immediately see the flow respond.</p> <h3> Example use case </h3> <p>A simple comparison like sedan vs hatchback geometry shows different wake structures:</p> <ul> <li>Sedans tend to have delayed separation and a narrower wake</li> <li>Hatchbacks separate earlier and form a larger recirculation region behind the rear surface</li> </ul> <p>These differences also relate to real effects like rear window soiling and why rear wipers are common on hatchbacks.</p> <h3> Important context </h3> <p>This is an experimental research framework, not a replacement for industrial CFD tools.</p> <p>Force coefficients like Cd and Cl are useful for showing trends inside the simulation, but they are not intended to represent fully validated engineering values.</p> <p>The focus is on:</p> <ul> <li>Flow intuition</li> <li>Real-time exploration</li> <li>Testing solver ideas</li> <li>Differentiable and interactive CFD workflows</li> </ul> <h3> What I’ll share next </h3> <ul> <li>Solver structure and numerics</li> <li>SDF-based geometry pipeline</li> <li>Performance design choices in JAX</li> <li>Real-time CFD tradeoffs</li> </ul> python machinelearning computerscience opensource