Beyond Faster Simulation: Hitting NASA's 2030 CFD Benchmark and the Infrastructure for Physics AI

AI has drastically reduced the cost of producing software code, transforming early prototypes into inexpensive "price checks" rather than lengthy debates. However, engineering teams are still limited by the high cost of simulations and the fragility of geometry workflows in hardware. In 2015, NASA's CFD Vision 2030 study issued a significant challenge: enabling a single engineer to run and interpret 10,000 simulations in 24 hours without manual intervention. This turns out to be a dual challenge: 1) harnessing enough computing power at a reasonable cost, and 2) generating CAD boundaries and body-fitting meshing that does not break under parametric variation or require human intervention.

In this talk, we will explain how nTop and CoreWeave surpassed NASA's 2030 milestone four years ahead of schedule. They completed 10,000 Large Eddy Simulation (LES) CFD runs in just 32.2 hours. This was done across a drone platform design space comprising 2,400 variants. By combining nTop's robust implicit geometry engine with a GPU-native Lattice Boltzmann (LBM) solver, we eliminated geometry failures and meshing steps entirely. A single engineer deployed a lights-out pipeline running across 280 CoreWeave RTX Pro 6000 GPUs that sustained near 100% compute utilization without any interventions during the run.

Finally, we will demonstrate how generating 14 TB of structured 3D flow data can transform CFD from a late-stage validation tool into the primary training engine for real-time physics AI surrogates operating in implicit latent space.