PICasso builds photonic circuits from natural language

PICasso gets the structure of a photonic circuit right in up to 92.7% of cases and its function in 52%. That distance between drawing something that passes the design rules and drawing something that also meets the optical specification is the most honest number in the paper published on 29 August on arXiv.

The work, PICasso: An AI-Enabled Design Framework for Autonomous Optimization of Silicon Photonic Devices, presents a framework to synthesise, verify and optimise photonic integrated circuits (PICs) from specifications written in natural language. Silicon photonics is the substrate of the optical transceivers that move traffic inside data centres, and its design flow remains considerably more artisanal than that of digital electronics.

How the pipeline works

PICasso does not ask an LLM to write the final layout. It chains generation in two hops, from natural language to YAML and from YAML to GDS, with PDK-aware knowledge injection (the foundry design kit), automated placement and routing, DRC/LVS validation and SAX-based photonic simulation. The model produces a declarative intermediate representation and the rest of the chain, which is deterministic, checks what it proposed.

PIC-Set, the benchmark

To measure all of this the authors publish PIC-Set, a set of 36 parameterised design tasks ranging from basic photonic primitives to multi-component circuits. On it they compare several state-of-the-art LLMs under a unified protocol, without the abstract naming which ones, using new metrics: structural and functional Spec@k, optimisation efficiency and robustness under perturbations.

The figures the paper reports: structural Spec@3 of up to 92.7% and functional Spec@3 of up to 52% on the highest complexity circuits, always above plain LLM generation without the surrounding scaffolding. On the optimisation side, mean insertion loss drops from 4.98 dB to 3.25 dB, that is 1.73 dB less, around a third.

Why 52% is the interesting number

A structural Spec@3 of 92.7% means that, in three attempts, a well-formed circuit almost always comes out: legal geometry, correct layers, PDK rules respected. A functional Spec@3 of 52% means that, in the hard cases, half of those well-formed circuits do not do what was asked. It is the same gap that shows up in code generation when you move from it compiles to it passes the tests, and it explains why the simulation step is not decorative in this kind of system.

What agent builders can take from this

The transferable lesson is not photonic. PICasso works because nobody lets the model emit the final artefact: it emits YAML, and that YAML goes through the PDK, through DRC/LVS and through a simulator before becoming GDS. It is the classic shape of a pipeline with external tools, the pattern implemented in the Claude ecosystem with MCP servers: the model proposes, the tool verifies and the result of the verification goes back into the loop. When the domain has an objective validator that is cheap to run, that design performs considerably better than asking the model to correct itself by reading its own output.

It is also worth looking at what the paper does not say. Thirty six tasks are few to talk about coverage, robustness under perturbations is measured inside the same benchmark being proposed, and nothing in the abstract suggests the human designer leaves the loop.

We find the work useful above all because of PIC-Set. A public benchmark that separates the structural from the functional forces the next automated design demo to be measured by what the circuit does, not by how clean the layout looks.