Generative placement

Qfactr is an early-stage company building physics-informed generative ML for photonic integrated circuit layout. The workspace you can use today — physical coordinates, explicit routing, PDK-aware parts, and geometry-driven loss — is the foundation that direction is built on. Generative placement is the part of that research direction we are most often asked about, so this page lays out the goal honestly without overstating where it stands.

If you want the problem this is meant to solve, start with why photonic layout is hard. This page assumes that context and focuses on the proposed approach.

The vision

The goal is an in-house generative model that places photonic components at scale: given a set of components and the connections they need, it proposes arrangements on the physical canvas. Rather than treating placement as an abstract packing problem, the model is meant to perform routing-aware layout exploration — proposing arrangements that respect the physical difficulty of routing waveguides through dense systems.

That qualifier is the whole point. In photonics, where you place a component largely decides whether it can be connected at acceptable loss. Waveguides cannot cross freely, bends cost optical power, and components are large relative to the connections between them. A placement that looks compact can be impossible to route cleanly, and a placement that routes well may waste area. The aim is a model that internalizes this coupling between placement and routing instead of leaving it for a human to discover by hand, one arrangement at a time.

Concretely, the intent is to let a designer hand the model a circuit — components, the pins that must connect, and the physical constraints they care about — and get back candidate layouts to evaluate, rather than starting every arrangement from a blank canvas. See the AI assistant for how Qfactr already operates over structured circuit state; generative placement extends that same grounded approach to whole-layout exploration.

Why physics-informed

"Physics-informed" is not decoration. A generic placement model can produce arrangements that are geometrically tidy but physically unroutable, because it has no notion that a bend radiates power or that a crossing costs loss. The premise here is that placement proposals should be constrained by real routing and loss physics, not by abstract placement objectives alone.

• Routing feasibility is a placement property. Two parts that are easy to wire in one arrangement may be impossible to connect at acceptable loss in another. The model is meant to reason about feasibility while it places, not after.
• Geometry is loss. Every bend, waveguide crossing, and detour costs optical power. Proposals should prefer arrangements whose routes can stay clean, consistent with how Qfactr already derives transmission and loss from real path geometry.
• Components have real footprints and pins. Placement operates on parts that carry physical pin positions and S-matrix data — the same PDK-aware parts the workspace uses — so a proposed arrangement is grounded in devices that behave like the ones a foundry can build.
• Proposals stay checkable. Because everything lives in the same physically grounded model, a proposed layout can be inspected, edited, routed, and simulated for loss like any other design — not accepted on faith.

Where it stands today

To be unambiguous: generative placement is in development. You cannot ask the workspace to place a circuit for you today, and the related goal of stronger automatic routing is likewise on the roadmap, not shipped. The capabilities that exist now are the manual, physically grounded ones — placing components at real micron coordinates, routing waveguides explicitly between pins, composing hierarchical blocks, and reading geometry-driven loss.

Both columns serve the same end: design photonic circuits so you can iterate in days, not weeks. Manual tools shorten the loop today; generative placement is meant to shorten it further by proposing strong starting points instead of requiring every layout to begin from scratch.

Next steps