q-twin · digital twin for photonic chips
Predict a photonic chip before it is made.
q-twin is a digital twin of your photonic chip. Describe it with your foundry’s PDK and what fabrication must deliver, run it as many virtual chips, and see which gates fail before you pay for a tape-out.
$ python -m studies.s2_calibration --design <pack>1Features
- Validate before fabrication: every acceptance gate checked on virtual chips at nominal, weak and best corners, before the first wafer
- Your layout stays private: describe the chip with parameters and expectations in 6 plain files; no GDS, no layout upload
- Built on foundry PDKs: choose an existing foundry PDK from the supported list; component models and tolerances come from it. CORNERSTONE coming soon
- A library of studies: loss, calibration yield, interference, gate fidelity, drift, sensitivity and design comparison, each scored against your gates
- Evidence on every number: measured, literature, calculated or target, and each result carries the worst label of its inputs
2Applications
- Pre-tape-out design validation
- Programmable photonic meshes
- Tolerance and yield budgets
- PDK and foundry choice
- Quantum photonic circuits
- Photonic AI and signal processing
3Description
A tape-out takes months and a mask set. q-twin builds one virtual chip per seed from your design pack, with loss, split-ratio spread, heater nonlinearity, thermal crosstalk, drift and detector imperfections, runs many of them, and reports the probability that the design passes each of its own gates.
One engine serves every design; only the pack changes. For one or two photons a linear optical network is computed exactly from its N × N matrix, so the uncertainty lives in the inputs, and the twin shows it instead of hiding it.
Table 1. Device Information
| Part | Package | Body size |
|---|---|---|
| QT-1 | design pack | 6 files (YAML) |
| Interface | web · CLI | PDF report |
| Studies | S-1 … S-8 | 3 corners each |
4Why a Digital Twin
A photonic tape-out costs a mask set, a fabrication slot and months of waiting. Most of what can go wrong is decided before the wafer exists. q-twin moves the first test before the first wafer, so a failing gate costs a rerun, not a fabrication run.
4.1Validate before you fabricate
The twin runs your design as many virtual chips at the nominal, weak and best corners of your process, and scores every acceptance gate. You see which gates bind, and by how much, while a change is still a line in a file.
4.2Your design stays yours
You describe what the chip is and what fabrication must deliver: mesh topology, component parameters, test bench and gates. No GDS, no layout file. Packs are data only, the twin runs no user code, and projects are private to their owner.
4.3Pick your foundry’s PDK
Choose an existing foundry PDK from the supported list. Component models and tolerances load from it, labelled with the PDK as their source. Generic specifications work today;CORNERSTONE support is coming soon.
Table 4-1. Where a design problem is found
| Without a twin | With q-twin | |
|---|---|---|
| When | After fabrication, on the test bench | Before tape-out, in a study run |
| What it costs | A new mask set and fabrication run | A change to the pack and a rerun |
| What you learn | One chip, one measurement | Pass probability over many virtual chips, at every corner |
| What you hand over to test it | The layout (GDS), to a simulation or test partner | Parameters and expectations only; no layout |
5Functional Diagram and Pack Map
q-twin models the chip you will tape out: light enters at the facet, crosses a mesh of interferometers and reaches the detectors. Every element is a line in the design pack. Hover or focus a number to see the line that drives it, or the result a run gives back.
# design.yamlmodes: 6mesh: { family: clements }cell: { splitter: mmi2x2, theta: undercut }io: { ports_in: [photon×6, reference, loopback×2] }# pdk.yamladapter: generic_specs# profile.yamlheaters.p_pi_mW: value: [100, 300] status: literature corners: { best: 100, weak: 300 }# bench.yamldetectors: { kind: apd, efficiency: 0.25 }# gates.yaml- { id: G4, name: calibration, threshold: "mean ε < 1 %" }
demo-6 v0.3 · 2026-11-12 · frozenS-2 50 virtual chips × 3 cornersG4 nominal PASS weak FAIL best PASSlabel target (worst input)fig fig_yield.svgreport report.pdfctl compare v0.2 → v0.3 · download pack
Sample pack and output for an illustrative design. Not a real chip.
Table 5-1. Pack Map
| No. | Name | I/O | Pack file / output | Description |
|---|---|---|---|---|
| 1 | TOPO | I | design.yaml mesh | The mesh: Clements, Reck, multiport or custom, and its number of optical modes. |
| 2 | CELL | I | design.yaml cell | One Mach–Zehnder interferometer: two couplers, two arms, and the phase shifters on them. |
| 3 | IO | I | design.yaml io | Edge couplers at the facet and their roles: photon, laser, reference, loopback, monitor. |
| 4 | PDK | I | pdk.yaml | Waveguide, bend, coupler and heater models from the PDK adapter, labelled literature with the PDK as source. |
| 5 | PROF | I | profile.yaml | Every number, such as heater Pπ, loss and split ratio, as value, status and source. |
| 6 | CORN | I | profile.yaml corners | Nominal, weak and best. A design range, like a coupler’s split spread, resolves by corner, never per element. |
| 7 | BENCH | I | bench.yaml | Off-chip: photon source, detectors, driver and temperature control. |
| 8 | GATE | I | gates.yaml | The design’s own acceptance gates, read on the outputs. Studies read them; nothing is hard-coded. |
| 9 | VER | O | run record | Each run reads a frozen version of the pack and records its version, date and commit. |
| 10 | FIG | O | figures | SVG figures per study: yield, loss, visibility and gate curves. |
| 11 | RPT | O | report | Report per run with inputs, method and verdicts. Download as PDF. |
| 12 | EVID | O | evidence label | Every result carries the worst evidence label of its inputs. |
| 13 | YLD | O | pass probability | Monte Carlo over virtual chips: the share of chips that pass each gate. |
| 14 | MTX | O | gate × corner | Every gate at every corner on one matrix, with verdict words, not colour alone. |
| 15 | CTL | I/O | CLI · dashboard | Run, compare versions, see what changed, download pack and report. |
6Study Run Timing
A study run is a signal you can read. The pack is frozen, loaded and checked, then each corner simulates its batch of virtual chips and is scored against the pack’s gates. Run it, then weaken the corner and watch the verdict change.
6.1Recommended Run Conditions
| Parameter | Min | Typ | Max | Unit |
|---|---|---|---|---|
| Virtual chips per corner | 1 | 50 | —‡ | chips |
| Target unitaries per chip | 1 | 100 | —‡ | targets |
| Corners per run | 1 | 3 | 3 | corners |
| Parallel workers | 1 | 4 | —‡ | workers |
6.2What each gate reads
| Metric | Study | Example gate |
|---|---|---|
| Fibre-to-fibre loss | S-1 | ≤ 4.0 dB |
| Calibration error | S-2 | mean ε < 1 % |
| HOM visibility | S-3 | > 95 % |
| CZ truth-table fidelity | S-4 | > 90 % |
| Phase drift | S-5 | < 5 mrad/min |
‡ Limited by your quota on the public host, or by your machine when self-hosted. Example gates come from an example pack; every pack sets its own.
7Typical Characteristics
The twin answers with curves, not single numbers: how yield falls as a parameter spreads, and how far each corner sits from the gate. Both graphs below are illustrative shapes, drawn to show what a study returns, not results for any chip.
8Detailed Description
Build, run, read. Each step works on the same pack, so a result can always be traced back to the exact version and values that produced it. Use the hosted platform in a browser, or the same engine on your own machine.
8.1Build
Build the pack on 7 screens: Basics, Topology, PDK, Parameters, Bench, Gates and Review, with the YAML shown live. Start blank, from an example, or import a .zip. The same loader as the command line flags any value without a status or source before you save.
$ ls designs/my-chip design.yaml pdk.yaml profile.yaml bench.yaml gates.yaml README.md 6 files · every value labelled
8.2Run
Queue a study on our server within your quota, or run it locally. Each run reads a frozen version of the pack. The jobs list shows what is queued and running, and a failed job keeps its log.
$ python -m studies.s2_calibration \ --design my-chip --corner all nom weak best 50 chips each G4 PASS FAIL PASS
8.3Read
The dashboard leads with what binds: the gates that fail or are at risk. Below it, the gate × corner matrix, version history, what changed between versions, and a PDF report. Verdicts are always words, never colour alone.
$ open dashboard · my-chip v3 binds G4 calibration · weak FAIL risk G1 loss · weak AT RISK report.pdf · worst label: target
Figure 8-1. Dashboard gate matrix, sample
| Gate | Requirement | Nominal | Weak | Best | Evidence |
|---|---|---|---|---|---|
| G1 loss · S-1 | ≤ 4.0 dB | PASS 3.1 dB | AT RISK 3.9 dB | PASS 2.6 dB | literature |
| G4 calibration · S-2 | mean ε < 1 % | PASS yield 93 % | FAIL yield 38 % | PASS yield 98 % | target |
| G6 HOM · S-3 | > 95 % | PASS 97.2 % | PASS 95.6 % | PASS 98.1 % | target |
| G8 drift · S-5 | < 5 mrad/min | NOT RUN queued | NOT RUN — | NOT RUN — | target |
9Typical Application
Replacing the per-chip pile. On the left, what a team typically keeps by hand to answer the same questions for one chip. On the right, what it keeps with q-twin.
Before: per chip, by hand
- 01Loss budget
spreadsheet - 02Calibration error
script per mesh size - 03Interference visibility
notebook - 04Gate fidelity
separate model - 05Drift interval
estimate in a doc - 06Sensitivity
one input at a time, by hand - 07Design comparison
table copied between docs - 08Where a number came from
a footnote, if any
After: one pack, every study
gates: - { id: G1, name: loss, study: S-1, threshold: "≤ 4.0 dB", status: target } - { id: G4, name: calibration, study: S-2, threshold: "mean ε < 1 %", status: target } - { id: G6, name: HOM, study: S-3, threshold: "> 95 %", status: target } # corners, labels, figures and reports# come with every run
Corners, evidence labels, figures and the report are properties of the run, not documents beside it.
Table 9-1. Example packs
| Example | Modes | What it shows |
|---|---|---|
| 4-mode test chip | 4 | loss, calibration and interference |
| 6-mode mesh | 6 | adds a post-selected two-photon gate |
| 16-mode mesh | 16 | adds an on-chip encoder for photonic AI |
3 example packs ship with q-twin, read-only. Copy one into a project to start, or begin blank with your own design. Their values are design targets, labelled as such.
Describe one pack. See every gate before tape-out.
Install the engine and the dashboard, copy an example pack, and run your first study. Every result arrives with its corner, its verdict and the evidence behind it.
$ pip install -e ".[api,test]"| Option | Status | Get it |
|---|---|---|
| Source (FSL-1.1-ALv2) | SOON | Source · soon |
| Self-hosting guide | PREVIEW | Get started |
| Hosted public beta | PREVIEW | Request access |
| CORNERSTONE | SOON | Get notified |
| Partner labs & foundries | ACTIVE | Talk to us |