At radio frequencies, layout is still design

The transition from schematic to layout is not simply the translation of a finished idea into coordinates. At radio frequencies, entering layout means continuing to design.

This distinction changes what I expect from an AI design system. Generating a plausible circuit is valuable, but it is not equivalent to delivering a working physical implementation. Nor does producing a layout file establish that the physical problem has been solved.

Connectivity does not determine behavior

One connectivity graph permits many geometries. The widths, spacing, layer choices, interconnect lengths, and surroundings of those geometries can change the electromagnetic response. Mutual coupling and parasitic effects are not always small corrections. They can be intentional parts of the circuit.

That is why layout-aware electromagnetic simulation belongs inside an RF design loop. For example, Keysight’s description of Momentum explicitly includes arbitrary multilayer shapes, coupling, and parasitic effects. The engineering implication I draw is simple: geometric freedom needs physical feedback.

Use the software analogy carefully

It is tempting to compare design-rule checking with syntax checking, layout generation with compilation, and simulation with testing. This analogy helps explain the workflow, but it also hides an important difference.

A compiler normally aims to preserve a program’s specified semantics. A physical RF implementation still contains consequential design choices that the original netlist may not uniquely specify. The “translation” can alter the behavior we care about. That means the loop must return to design decisions, not merely repair formatting errors.

Separate three kinds of evidence

I find it useful to distinguish manufacturing consistency, connectivity or interface consistency, and functional performance. They answer different questions. Passing one should not be presented as passing the other two.

Conventional layout-versus-schematic checks remain useful where device identity and connectivity are part of the contract. But there are RF structures whose distributed behavior is the object being designed. For those structures, I want explicit ports, reference planes, boundary conditions, and response requirements rather than an arbitrary insistence that every shape resemble a preconceived lumped implementation.

This is not an argument to remove consistency checks. It is an argument to make the check match the abstraction. A passive electromagnetic region still has interfaces and constraints; they simply need to be specified at the right level.

A hierarchy should permit iteration

Floorplanning, component geometry, routing, and matching are coupled. I do not want a workflow that refuses to plan the floorplan until every component has a finalized performance model, or one that fixes the floorplan so early that subsequent physical learning becomes unusable.

A more useful contract starts with provisional envelopes: ports, rough bounds, keep-outs, intended roles, and uncertainty. As geometry and simulation become available, those envelopes are revised. An early placeholder is acceptable if it is labeled as such and cannot silently pass as a verified component.

The agent’s role is to manage those dependencies and decide which revision is worth making. The tool layer should support inspection and localized edits, without pretending that a generic placement heuristic knows the RF objective.

Deliver something that can be continued

The final handoff should preserve more than a screenshot. I care about editable design state, the geometry that was actually evaluated, the setup used to evaluate it, and a record of remaining limitations. A beautiful rendering with uncertain provenance is weak evidence.

Autonomous RFIC design becomes interesting when the system can navigate this circuit–physical loop, respond to inconvenient feedback, and leave behind an implementation another engineer can understand and modify. The hard part is not drawing the layout. It is closing the right loop.

Technical reference

• Keysight. Momentum Key Features. Background on multilayer geometry, electromagnetic coupling, and parasitic effects. The workflow recommendations above are my own.