Physics AI: the 3D data bottleneck

Physics AI is moving fast. Neural surrogates that predict drag, lift or stress in milliseconds instead of hours, generative models that propose shapes for a target performance, foundation models trained across whole families of simulations: the promise is real, and the research community is racing towards it. But there is one ingredient that every one of these models needs, and that the field is chronically short of: good 3D geometry, in large quantities.

Text models had the internet. Image models had billions of photos. Physics AI has… a handful of curated datasets, each of which took a heroic effort to assemble. The bottleneck isn't the architectures or the GPUs. It's the data, and more specifically the geometry at the start of the pipeline.

Why 3D data is so hard to produce

To train a model that learns physics from shape, you need many shapes. Not a few hand-picked examples, but hundreds or thousands of variants spread across a meaningful design space, each one clean enough to be meshed and simulated without anybody touching it.

That last requirement is where traditional CAD breaks down. As we discussed in the quest for the right geometry kernel, classical B-rep kernels are a patchwork of separately-defined surfaces stitched together, and nothing guarantees the stitching holds. Push a parametric model through a few hundred parameter combinations and some of them will fail: a boolean that doesn't resolve, a fillet that collapses, a sliver face, a model that isn't watertight. At the scale of a dataset, "most of them work" isn't good enough. Every broken sample needs a human to repair it, or worse, it silently pollutes the training set.

So the field has found a workaround: morphing. You take one carefully prepared baseline geometry and deform it. The best example is DrivAerML, an outstanding open dataset of 500 variants of the DrivAer car, each simulated with high-fidelity hybrid RANS-LES CFD. The 500 geometries were generated by placing morphing boxes around the baseline car and deforming it along 16 parameters (length, width, height, windscreen angle, and so on).

DrivAerML is a landmark, and it has rightly become a reference for the community. But morphing has limits that are built into the method:

• The topology never changes. Every sample is the same mesh, stretched. You can't add a blade, remove a hole or change the number of anything.
• The parameters describe the deformation, not the design. A morphing box is a proxy for engineering intent, not the intent itself.
• The design space is anchored to one baseline. You explore the neighbourhood of a single shape, not a family of designs.

Parametric, watertight, at scale

This is the gap the Kioko kernel was built to fill. Because it represents a body as a single watertight surface rather than a stitched patchwork, a Kioko model doesn't fall apart when you sweep its parameters. Every variant comes out as a closed, valid solid, ready to mesh and simulate, with no repair step and no human in the loop.

That changes how a dataset can be built. Instead of morphing one shape, you write the design once, as a parametric model driven by real engineering inputs, and let the kernel generate as many variants as you need. The parameters aren't deformation handles. They are the actual design variables an engineer would reason about, and they can change the topology as freely as the dimensions.

Watertight datasets of complex geometries are now available through our API. You define the design space, and we generate the geometry.

A first dataset: Kioko Propellers

To show what this looks like in practice, we are releasing our first open dataset on Hugging Face: kioko-propellers. The spinning grid at the top of this post is a sample of it: every propeller there is a distinct design, and each one is a watertight solid generated by the Kioko kernel.

The dataset contains 500 propellers, each generated from a single parametric model:

• Blade count from 4 to 6. This is a topological change that morphing simply can't express.
• Span, chord distribution and blade offsets defined at several stations along the blade.
• Section-level aerodynamic parameters (angle of attack, leading-edge and trailing-edge shape, thickness) at multiple stations from root to tip.

Each sample ships with its full parameter vector, a watertight STL mesh and front and back renders, under an MIT licence. The parameters and the geometry are paired one-to-one, so the dataset can be used both ways: to learn shape from parameters, or parameters from shape.

Following in the footsteps of DrivAerML

Our ambition is simple. We want to follow in the footsteps of DrivAerML and help build the open data foundations that physics AI needs, with one difference in method: datasets driven by parametric inputs rather than morphing. That means design spaces that mean something to an engineer, topology that is allowed to change, and geometry that is guaranteed to be valid every single time.

The propellers are a first step. We will keep releasing open-source datasets on Hugging Face to help the research community develop the next generation of tools, and we'd like the community to help shape what comes next.

We'd love your feedback

This is where we need you. If you work on physics AI, geometric deep learning, surrogate modelling or design optimisation, we would love to hear:

• What you think of the current dataset. Is the design space useful? Are the parameter ranges sensible? Which formats, resolutions or extra outputs (simulation results, point clouds, SDFs, B-rep) would make it more valuable for your work?
• Which dataset we should create next. Which family of geometries would unblock your research? Heat exchangers, brackets, wings, turbomachinery, lattices, something we haven't thought of?

You can open a discussion directly on the dataset page, or write to us at contact@kioko.io. And if you need a watertight dataset of your own, built around your own design space, the API is ready.