Well done! 👏
Turbulent flow around a car body, simulated with a solver that also shows how design changes affect the result.
This video presents large-eddy simulations of the Ahmed body, a reference geometry in road-vehicle aerodynamics, at a Reynolds number of 768,000 and rear slant angles of 25° and 35°. It walks through the measured geometry and its representation in the solver, the 94-million-cell grid, and the governing equations and turbulence model. It then explores the transient vortex structures and wake, before comparing the results with laser-Doppler measurements.
What sets the framework apart is the combination of experimental validation, full differentiability and efficient GPU performance.
Reverse-mode differentiation runs through the entire solver, including the pressure projection and the immersed body. The resulting gradients agree with finite differences within the precision those checks can resolve. A single backward pass gives the sensitivity of drag to the slant angle.
The solver is built for GPUs and runs in double precision, from a single GPU to multi-node jobs on the NVIDIA GH200 Grace Hopper nodes of the Alps supercomputer @cscsch. On one GPU, it achieves around half the peak memory bandwidth. In scaling tests, a billion-cell grid advances by one time step in 78 milliseconds on 32 GPUs.
These capabilities matter because foundation and world models for physics are only as good as the data they learn from. They need reference simulations of known accuracy and, increasingly, gradients alongside snapshots to learn how a flow responds to changes in shape or operating conditions. A validated, differentiable solver that runs efficiently on modern GPUs can provide exactly this kind of training and test data.
The analysis also offers a sobering reminder of the challenge: small errors in this flow grow a millionfold within roughly twice the time it takes the flow to travel the length of the body. A learned model of this wake should therefore be judged by its statistics and its response to design changes, rather than by how closely it follows a single trajectory.
More is coming. Stay tuned.
#DifferentiableSimulation #FoundationModelsView quoted post on X
