There is a version of materials science that looks less like chemistry and more like query planning: define the behaviour you need, search the space of structures that could produce it, and manufacture the survivor.
Structure has always been the lever — lattices, laminates, foams, the way a bone is dense only where it must be. What changed is that the search space is now large enough to be interesting and cheap enough to explore. A twin costs nothing to destroy, so we destroy thousands.
Loop, not pipeline
Our workflow is embarrassingly simple in outline. Specify the load case. Generate candidate geometries. Simulate to failure. Keep the survivors and mutate them. The difficulty is entirely in the fidelity question: how wrong is the simulation allowed to be before the loop starts optimising for the error rather than the world.
Every simulation is a hypothesis with a manufacturing bill attached.
We calibrate against physical tests more often than is convenient. Roughly one in forty candidates gets built. The build is where the model learns what it was quietly lying about.
Why we care
Spatial systems need bodies: enclosures that dissipate heat without fans, surfaces that hold sensors without looking like sensors, structures that are stiff at one frequency and dead at another. None of that is available off a shelf. So the materials track and the perception track are the same programme, viewed from two ends.
We expect the first public result of this work to be a component, not a headline. That suits us.