Beyond Faster Simulation: Hitting NASA's 2030 CFD Benchmark and the Infrastructure for Physics AI
nTop and CoreWeave completed 10,000 LES CFD simulations in about 32 hours, meeting NASA's 2030 CFD Vision goal four years early. The approach combines signed distance field geometry, which cannot break under parametric variation, with a GPU-native Lattice Boltzmann solver that removes the meshing step entirely. The talk also covers the 14 TB dataset generated and the path toward physics AI surrogates trained in implicit latent space.
Max, representing nTop, walks through a proof-of-concept study run with CoreWeave that executed 10,000 large eddy simulation (LES) CFD runs as a preliminary step toward physical AI development. The motivation is NASA’s CFD Vision 2030 study from 2015, which set the goal that “a single engineer must be able to conceive, create, analyze, and interpret a large ensemble of related CFD simulations in a time critical period without individually managing each simulation.” The speaker argues this is still far from reality, and the gap comes down to two limiting factors: fragile geometry pipelines and meshing, and the compute cost and scalability of solvers.
From software agents to hardware iteration
Max opens with an analogy to software. Large language models have sharply reduced the cost of producing code: an agent can take an intent and generate, test, deploy, and maintain software that follows it. Hardware development has no equivalent yet. The vision presented is to bring that same iteration loop to physical products. Today much CFD work consists of babysitting convergence criteria and residuals on multi-day jobs, and repairing meshes and numerical parameters by hand. The NASA challenge aims to remove that manual layer so engineers can focus on interpreting results.
Why the challenge is hard
The design space has two coupled parts: a geometric parameter space and the physical performance of each variant. Only the combination yields engineering insight, such as identifying the best compromise between weight, manufacturability, and pressure drop or flow performance.
Traditional CAD based on boundary representation (B-rep) does not scale to thousands of variants automatically. The speaker identifies three kinds of fragility:
- Geometric fragility: tangencies fail to line up because the generated surfaces are not mathematically well defined.
- Topological fragility: large parametric changes alter topology, so features must be recreated, and healing tools are manual.
- Reference fragility: edges and surfaces exist or disappear depending on parameter values, so automated referencing is error-prone.
Meshing is the second unsolved problem. After more than 30 years of CFD, producing a mesh that is both quality-controlled and accurate for the solver remains manual and iterative. Across thousands of geometries, failures appear as self-intersecting or overlapping faces, or as solvers struggling with meshes that are technically valid. Finally, many solvers still run on thousands of CPUs for weeks, which makes large studies expensive.
Implicit geometry with signed distance fields
The first pillar of the approach is unbreakable geometry based on signed distance fields (SDFs). Instead of describing a shape by its boundaries, the geometry is a scalar field giving the signed distance to the surface: negative inside the body, positive outside, and zero exactly on the surface. There are no explicit edges, vertices, or face references, so there is nothing to break. Booleans, rotations, and blends reduce to basic field math that parallelizes efficiently on GPUs. Using this approach, the team synthesized 2,400 drone geometries across six geometric parameters without a single failure.
nTop encodes these SDFs much like an abstract syntax tree, which allows it to represent an entire parametric design space rather than a single geometry. Dependencies between inputs can be expressed directly. For example, changing a wing shape automatically propagates to the internal lattice structure.
A Lattice Boltzmann solver that bypasses meshing
The second pillar is nTop Fluids, a GPU-native Lattice Boltzmann Method (LBM) solver. LBM is voxel-based, so it needs no conformal mesh. Voxelization follows directly from the SDF by sampling its sign to produce a voxel mask. The algorithm’s structure is highly parallel, which suits GPUs well. Because nTop Fluids is built as a block within nTop’s own element structure, every geometry update automatically triggers a new simulation. The speaker contrasts this with collections of scripts that couple geometry generation, solving, and result back-propagation, describing it instead as one unified design workflow.
On turbulence, LBM is inherently transient, so the team pairs it with LES, which resolves the large turbulent structures and models only the smallest scales. This differs from the industry-standard RANS approach, which models all turbulent structures and their time variance.
Compute and execution on CoreWeave
CoreWeave provided 280 NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. Each simulation runs on a single GPU in isolation with no GPU-to-GPU communication, making the design-of-experiments workload embarrassingly parallel. Over the run, the system maintained near 100% computational saturation for just under 40 hours.
The study covered a group 3 long-endurance fixed-wing UAS drone. Six geometric parameters, including length, span, inboard and outboard trailing edge angles, were sampled using Latin hypercube sampling. The 2,400 geometries were each simulated at five angles of attack, for a total of 12,000 simulations. Progress was cumulative: about 7,500 jobs were complete after 24 hours, 10,000 after roughly 32 hours, and all 12,000 just short of 39 hours. The workflow was orchestrated with nTop Orchestrate, which configures the design of experiment and dispatches jobs to Kubernetes for queuing.
Results and dataset
Beyond execution, the dataset yields performance insight. The speaker analyzed lift and drag coefficients and notes that the dataset needed this size to detect smaller influences on performance. A Pareto front between aerodynamic efficiency and structural weight illustrates the central trade-off: the two cannot be maximized together, so making a sound design decision requires identifying the front. The authors frame this as multidisciplinary optimization, enabled by a scalable approach that traditional CAD and CFD could not support. The dataset totals about 14 TB of structured 3D flow data and is available on request through the nTop community.
Toward surrogate models in latent space
The next step is training surrogate models on this data, and several partners are already exploring it. The longer-term goal is to close the loop by training surrogates directly on implicit geometry in latent space, so that no boundary representation or mesh is needed at any stage. This would allow scalable surface or volumetric field predictions in near real time. In that picture, CFD becomes the primary training engine for physics AI rather than a late-stage validation tool.
Conclusion
The core message is that agent-driven, software-like iteration is becoming possible for physical hardware. By combining unbreakable implicit geometry with a CFD pipeline that eliminates conformal meshing, the team met NASA’s 2030 benchmark four years early. The speaker hopes engineers and agents will stop fixing broken meshes and instead direct autonomous AI across massive design spaces.