Cogitan
← All surrogates

Fluxus-Q

Served with caveats

Radiation-correlated errors

For a proposed qubit chip: how often cosmic-ray muons and environmental gammas cause correlated error bursts, how many qubits a burst covers, how much back-side phonon traps would help, and where a correlated floor starts to bind for a given code distance.

For: QPU teams without an in-house quasiparticle or radiation group who need to weigh die thickness, back-side traps, shielding and siting before fabrication.

9.5/hr

muon burst rate vs Willow's measured 9.3/hr

only at an assumed 0.3 mm die · 3.84× high at 0.5 mm

Status. Served with seven standing known misses, returned on every response.

Engine. A calibrated physics chain, graded against G4CMP phonon simulation and compared with published device measurements.

Reference. Comparisons to published device measurements — not predictions of your fabricated hardware — plus the G4CMP phonon simulator for the phonon-spreading step.

API

Endpoints

POST/v1/fluxus/rad/assess

one design → rate, footprint, trap correction, floor

POST/v1/fluxus/rad/assess/batch

1–25 designs, within a work budget

Base URL https://api.cogitan.ai, bearer key from an approved account. Read GET /v1/fluxus/capabilities first: it states the same fidelity, validation, envelope and refusal conditions as this page. Access model.

Inputs

What you send

die_mm
mm — die edge length
substrate
silicon | sapphire
thickness_um
µm — snaps to the nearest cached thickness, and says so
n_qubits
pitch_mm
mm — qubit pitch
film_coverage
fraction
backside_cu_coverage
fraction — back-side trap coverage
trap_pitch_mm
mm
trap_absorption
fraction
lead_cm
cm — lead shielding
underground
true | false
source
muon | gamma-lines | gamma-flat
detection_threshold_keV
keV — a device property; no default
code_distance
optional
lam
error-suppression factor Λ for the crossover distance
n_events
20 – 3,000 — simulated strikes per assessment

Outputs

What comes back

burst_rate
/hr — with the muon and gamma split; gamma served as a band
footprint_qubits
burst size in qubits; carries a stated +25% systematic
trap_correction
effect of back-side traps, ±9.2% 90% band
logical_floor
a published measured floor used as a profile — not ours, labelled NOT_OURS
calibration
per section: CALIBRATED, BORROWED or REFUSED
refusals, known_misses
always present

Validation

Measured against what, and how

Reference. Comparisons to published device measurements — not predictions of your fabricated hardware — plus the G4CMP phonon simulator for the phonon-spreading step.

Split. Not a train/test split: each figure is a comparison against a named published measurement, or against G4CMP at a named design point.

Muon burst rate, Willow

9.5/hr vs 9.3 measured

only at an assumed 0.3 mm die; 3.84× high at 0.5 mm, 6.03× at 0.725 mm

Gamma burst rate, Willow

9.8–13.5/hr band

against 41.4/hr measured; point estimate 0.29× at 0.3 mm

Median burst size, Willow

14 vs 15 qubits

biased ~25% high (see known misses); p90 withdrawn

Burst duration, Willow

2.03 ms vs 1–2 ms

Muon detection efficiency, Li et al.

100% vs 99.5%

a weak test — it saturates

Phonon spreading vs G4CMP

10.2% misplaced

against a 6.64% statistical noise floor

Trap correction

±9.2%

90% band

Speed

What it costs per call

One assessment

0.55 s + 0.0033 s / event

vs G4CMP

~22,000×

at a design point with traps present

Envelope

Where the answer is trusted

die thickness
cached at 0.3 / 0.5 / 0.725 mm — other values snap to the nearest and say so
trap_pitch_mm
0.1 – 2.0 mm
backside_cu_coverage
0.05 – 0.75
trap_absorption
0.2 – 1.0
graded points
burst rate: two published devices; footprint: one — anywhere else a section is labelled BORROWED
batch
1–25 designs, ≤ 60 reference-seconds of work

Refusals

When you get no number

  • No detection_threshold_keV → no burst rate. It is a property of the device, so it is refused rather than invented.
  • Trap geometry outside the trained box → no trap correction, with the reason.
  • Away from a graded point a section is labelled BORROWED and names the axis it departs on and the device whose normalisation it reuses.
  • A batch over the work budget is refused up front with the size that would fit.

Known limits

What it does not do

  • Geometry is not uniquely identified: the transport thickness and effective film absorption were fitted together to burst size, and other pairs fit equally well.
  • The gamma flux is inferred from a second device in a different building and applied to the first. A factor of a few on the gamma rate is expected.
  • No single die thickness fits both channels. The muon rate is right only on a thin die (0.3 mm) and the gamma rate only on a thick one; Willow's thickness is not published.
  • Phonon transport in the served chain is not G4CMP. G4CMP graded the spreading and fitted the trap correction, nothing more.
  • The per-qubit response normalisation is fitted to Willow and borrowed on any other device.
  • Burst footprints carry a +25% systematic: against G4CMP's own spreading, burst sizes are biased high by about 25%, and the 90th-percentile size is not a validated quantity.
  • Non-ionizing bursts are not modelled. Radiation is the ~17–18% minority of bursts; in a well-shielded device the non-radiative source dominates.

Where something else wins

  • For one certification-grade answer at a single design, run G4CMP. This is for ranking mitigations across many designs, where relative ordering is what matters.
  • The logical floor is a published measured floor (arXiv:2408.13687), served as a profile. It is not our prediction, it is not attributed to radiation, and the response says so.

Every figure on this page compares against a simulator or a published device measurement. None of it is a prediction of your fabricated hardware. For one certification-grade answer, run the full reference; this is for screening many designs cheaply and spending solver time on the survivors.