SuperconducTED
A noise model fitted to one day of IBM calibration data can be stale by the next. SuperconducTED asks whether fuzzy inference can make Qiskit Aer models last.
Last updated: October 1, 2026
Status · Active research
Role · Lead & Architect
Can a noise model survive recalibration?
Quantum hardware is noisy, and that noise drifts. In the project's archive of calibration data from IBM's ibm_fez, a qubit's readout error changes about every four hours and its T1 and T2 about once a day (issue 45). A noise model fitted to one day's data can therefore be stale by the next.
SuperconducTED asks whether a model can capture noise in a way that transfers across calibration cycles. The approach uses fuzzy inference to turn IBM calibration data into ensemble-based noise models for Qiskit Aer simulations of IBM backends.
The Problem
Calibration-based noise models, including those current Qiskit Aer workflows build, describe one device state at one moment. The project's architecture document states its premise: such models match hardware tightly on the snapshot they were tuned on and degrade silently when calibration drifts. The research target is a noise representation whose predictions stay useful as the hardware drifts.
The Approach · Why Fuzzy
SuperconducTED fuzzifies three calibration features, the across-qubit means of T1, T2 and readout error, instead of treating them as exact scalars. It uses Takagi-Sugeno-Kang (TSK) inference, and both type-1 and interval type-2 fuzzy sets are implemented. Since July 2026 the team has preferred interval type-2 provisionally, pending a membership-function ablation (ADR-009).
The intuition: a type-1 system says the membership of T1 = 142 µs is 0.85; a type-2 system says it is somewhere in [0.55, 0.85], because we are also unsure how unsure to be. T1 is a property of the noise; type-2 is how much uncertainty we are allowed to carry about it.
Fuzzy inference sits in front of the simulator and complements the Kraus formalism. It maps the calibration features to amplitude- and phase-damping rates, which become Kraus channels that Qiskit Aer applies to single-qubit gates.
The Pipeline
A seven-stage pipeline turns one calibration snapshot into Qiskit Aer noise channels: calibration ingestion, feature extraction, fuzzification and TSK rule firing, defuzzification, squashing, gate eligibility and channel projection (README). Gate eligibility joined in September 2026 and stops noise landing on zero-duration rz gates, as the post on that change explains. Each stage is documented and version-controlled, with design choices recorded as architecture decision records.
The Calibration-Data Engine
Transferability can only be tested against real drift, so the project collects it continuously. Since 13 May 2026 a scheduled GitHub Actions workflow has polled IBM's ibm_fez backend and committed each new calibration snapshot to a dedicated calibration-data branch. Since September 2026 it fires four times an hour, because GitHub drops scheduled runs under load. A planned time-split test will use this archive to check whether a model fitted on one window still predicts the next.
Where the research stands in October 2026
There is no transferability result yet. The pipeline runs end to end, from a calibration snapshot to Qiskit Aer noise channels. Its membership functions are not trained yet, its ensemble members are still identical and the type-2 defuzzifier collapses each interval to one crisp value, so the engine does not yet produce an uncertainty interval. A paper is planned; its critical path includes the trainer, the interval output and the time-split test.
Team
A five-person team plus a faculty advisor at TED University. Mert leads the team, wrote the architecture document and keeps the decision record.
Tech Stack
Qiskit · Qiskit Aer · IBM Quantum · NumPy · pytest · GitHub Actions · Python
Links
GitHub organization → github.com/SuperconducTED
Noise engine source → github.com/SuperconducTED/superconducted-noise-engine