nTop and CoreWeave Demonstrate NASA's 2030 Grand Challenge in CFD is Within Reach Today

In 2015, NASA published its CFD Vision 2030 Study. One of the grand challenge targets: an engineer/scientist must be able to generate, analyze, and interpret a large ensemble of related simulations in a time-critical period (e.g., 24 hours), without individually managing each simulation, to a pre-specified level of accuracy and confidence.

The target date was 2030, and in April 2026, we exceeded major milestones.

nTop and CoreWeave ran 2,400 drone planform geometry variants across five angles of attack for a total of 12,000 simulations. One engineer set it up and walked away while 280 NVIDIA GPUs on CoreWeave were fully saturated for the entire run without any geometry failures.

Geometry Bottleneck in Lights-Out Workflows

When people talk about large-scale design space exploration, they usually talk about simulation speed. More GPUs, faster solvers, cheaper cloud time. Those matter, but they're not the only thing holding teams back.

Producing geometry at scale is also a very real bottleneck.

Traditional CAD and CAE workflows often break under variation: parametric updates fail, fillets overlap, and meshes frequently don't generate cleanly. Geometry and meshing failures in automated parametric studies are a well-documented limitation of traditional engineering tools. Failure rates scale with model complexity and parameter range. Every failure means a human has to stop, diagnose, and fix the model before the pipeline can continue. That results in a structural ceiling rather than a performance limitation.

The number of GPUs you have available doesn't matter when your geometry breaks at case 47. When that happens, someone has to manually intervene to diagnose and fix the problem before the pipeline can continue, which means you can't run 12,000 cases overnight. Compute capacity sits underutilized, and the pace of simulation work is ultimately limited by your fastest CAD modeler. That ceiling is what made the NASA 2030 target aspirational in the first place.

The Pipeline

The study used a fully parametric Group 3 long-endurance fixed-wing UAS model built in nTop. Six geometry dimensions covering root chord, inboard and outboard leading edge sweep, inboard and outboard trailing edge sweep, and spanwise panel break were evaluated across five angles of attack per variant. 12,000 samples were drawn from that design space. Each sample ran a full LES CFD simulation via nTop Fluids and returned structured results: surface pressure, force coefficients, L/D across the AoA sweep.

Zoom image

The pipeline ran headless — no geometry failures, no manual restarts, no one watching it.

CoreWeave supplied the infrastructure: 280 NVIDIA RTX Pro 6000 Blackwell GPUs, each with 96 GB of VRAM, enabling one full LES simulation to run concurrently per GPU. The cluster stayed fully utilized from submission to completion. The project was supported throughout by Justin Hodges, CoreWeave’s Head of Physical AI, contributing physics consulting expertise alongside the infrastructure — helping shape the experimental design and interpret the results.

Max Gaedtke, the nTop engineer who set up and ran the pipeline, described it as the largest nTop Fluids dataset ever generated.

Why the Geometry Held

nTop generates models using signed distance fields rather than boundary representations, which store geometry as a set of surfaces, edges, and vertices with explicit topological relationships between them. In nTop, parametric changes propagate through a mathematical function so there are no topological elements to break. A boolean union in traditional CAD is a face-face intersection operation that fails when surfaces become tangent or near-tangent. In nTop, it's min(F(A), F(B)) — an operation with no topological elements to break.

That's what makes autonomous execution possible. The pipeline does not need to be managed because every one of the 2,400 variants generated a valid geometry. Where jobs failed, the failures were solver-side and recoverable, not geometry failures, and everything else ran to completion.

The CFD Solver

Traditional CFD requires a body-fitted mesh — a volume discretization that conforms to the surface of the geometry being analyzed. Generating that mesh is a manual, iterative process when geometry changes and requires a trained analyst to repair before the next simulation can run. In a parametric study of 2,400 variants, that step would make autonomous execution impossible.

nTop Fluids uses a Lattice Boltzmann Method solver that operates directly on the implicit geometry representation. Because LBM discretizes the fluid domain on a fixed Cartesian lattice rather than a surface-conforming mesh, there is no body-fitted mesh to generate or repair. Geometry changes propagate through the notebook, and the solver reads the updated signed distance field without any additional preparation step.

Boundary conditions, flow parameters, and output definitions are specified once inside the nTop notebook and remain fixed across every variant. Each simulation in the study ran from the same setup, with only the geometry changing. The results — surface pressure distributions, force coefficients, and lift-to-drag ratios across each angle of attack — were written directly to structured output files for downstream analysis. No analyst in the loop, no per-case setup, no checkpoint required.

Infrastructure Fit

CoreWeave is building toward something specific: on-demand GPU compute with serverless-style abstractions, where engineers get the scale they need without managing infrastructure. The challenge with that vision has always been the workloads. Most engineering processes still require human intervention somewhere in the loop, whether it is geometry prep, mesh cleanup, job monitoring, and/or failure recovery. You can't build a fully autonomous GPU pipeline around a workflow that breaks and needs fixing.

nTop geometry generation is autonomous and scalable, enabling high-utilization. It doesn't need supervision to run at scale and – when used for large scale studies – benefits greatly from parallel compute. nTop produces geometry that doesn't need human intervention to run at scale. CoreWeave provides the compute infrastructure to absorb that at whatever scale the study requires. Together, they remove the friction point between a parametric model and a fully autonomous pipeline.

What the Study Produced

The study produced a Pareto front across lift, drag, and L/D that would be practically impossible to build through a manual DOE. But the more durable output is the dataset itself. 12,000 CFD results, each with full 3D flow fields containing surface pressure, volumetric data, and force coefficients across the AoA sweep. The result is more than answers to this specific problem. The geometry diversity and LES spatial resolution make it a viable training dataset for surrogate models that can predict aerodynamic performance across the planform space without re-running simulation.

Zoom image

So the pipeline doesn't just produce insights for just the initial program, it produces usable data to accelerate the next one.

What This Enables

The same approach applies to any parametric geometry space: inlets, heat exchangers, structural panels, propulsion geometry. The physics change with each application but the pipeline doesn't. A single engineer, a parametric model that doesn't break, and on-demand compute at whatever scale the study requires.

The NASA 2030 target isn’t only about compute speed. It is also about removing manual effort from the loop to make simulation at scale something a single engineer could set up, walk away from, and come back to with results. That's what was demonstrated here, years ahead of schedule.