Research preview · TED University, Ankara
Put a range on your quantum noise model.
A noise model built from one calibration is a single draw from a chip whose noise keeps moving. SuperconducTED is building a noise engine for Qiskit Aer that learns how far that noise spreads across the whole calibration history, so your simulation can report a range.
See how the engine works Install it and run the tests
Work in progress: the pipeline runs end to end, while training and the range output are still being built. Unless a measured result is cited, this site describes the aim.
A real qubit, drifting
Each dot is one real T1 calibration of qubit 77 on IBM's ibm_fez, May 8, 2026 to October 15, 2026. The dashed line is what a snapshot from the first day keeps assuming. The shaded band is an illustration (the spread of the previous 7 days), not engine output.
01 / The problem
Why one calibration misleads your simulation
Three steps take you from the hardware to your results. The chart above is the evidence for the first.
The chip drifts
T1, T2 and error rates change from one calibration to the next. Our archive holds 868 NC-069 distinct device states of IBM’s ibm_fez since May 2026.
Your model freezes one moment
NoiseModel.from_backendcopies the latest calibration into a fixed noise model. It is exact for that moment and is never updated.Your results inherit the gap
Run the circuit a week later and the model still uses last week’s numbers. Nothing in the output tells you how far off it is.
02 / The approach
A noise engine that learns from every calibration
SuperconducTED sits between IBM’s calibration data and Qiskit Aer, the simulator.
Learns how calibrations map to noise
A fuzzy rule base turns three features of each snapshot (mean T1, mean T2, mean readout error) into noise parameters; ANFIS training, now in progress, fits it to the archive.
Keeps its reasoning readable
Each rule covers one region of calibration space, such as low T1 with high readout error, so you can inspect what it learned. The rule base has 234 NC-045 trainable parameters.
Reports a range
The plan: an ensemble of noise models (32 NC-018 by default) turns one simulation into an interval of outcomes.
03 / Where it stands
The engine in October 2026
Each claim carries its evidence. Open a chip to see the record behind it.
Working
- The calibration archive of ibm_fez, polled since May 2026, with feature extraction.
- The fuzzy pipeline, end to end: from a snapshot to Aer noise channels.
- 859 NC-021 tests in the engine’s suite.
In progress
Until those decisions land, the engine’s outputs are point estimates.
Measured, with sources
156
qubits on IBM's ibm_fez, the machine the engine follows
NC-067 · as of Oct 16, 2026
868 of 1,170
distinct device states archived, toward the training floor
2,050
calibration snapshots archived since May 2026
NC-068 · as of Oct 16, 2026
859
tests in the engine's test suite
NC-021 · as of Oct 16, 2026
Every figure comes from the engine's numerical-claims register. Open a row id to check it.
04 / Questions
Before you try it
Can I use it in my research today?
You can install it and run the full pipeline and its test suite. Its outputs are still point estimates, so don’t cite them as results until training and the range output land. The Updates page will say when.
Do I need an IBM Quantum account?
Not to install the engine or run its tests. You need an IBM Quantum token only to poll live calibrations yourself. The archive the engine trains on is in the engine’s public repository.
Isn't rebuilding the model after each calibration enough?
A fresh model is a good single estimate, and we don’t claim it is wrong on average. What it can’t tell you is how far the hardware may sit from it. IBM recalibrates parts of a chip whenever its monitoring detects drift (IBM’s calibration guide), and a single routine T1 measurement predicts a qubit’s long-run average only weakly (Carroll et al. 2022). The engine is designed to learn that spread from the calibration history and report it as a range. We plan to judge the range by how often it covers repeated runs on the hardware, with the fresh model as the baseline.
Why fuzzy logic?
A fuzzy rule base stays readable: each rule names the calibration conditions it covers. Interval type-2 fuzzy sets can also carry the uncertainty of their own inputs, which is what a range needs. ANFIS trains the rules the way a small neural network is trained.
Is this an IBM project?
No. SuperconducTED is an independent student project at TED University in Ankara. It uses IBM’s published calibration data and the open-source Qiskit Aer simulator.
How do I cite it?
Use Cite this repository on the engine’s GitHub page. It reads the project’s citation file, which lists the authors.
05 / Updates
Follow the build
Get updates
One email per update
Almost done. If this address isn't subscribed yet, a confirmation link is on its way to your inbox. Click it to finish.
Next step
Run the engine on your own machine
You need Python 3.11 or newer and git. Install it, run the tests, then read the design.