Online anomaly detection in quantum device calibration streams

---

1. Problem

Superconducting and other quantum processors require frequent calibration to maintain acceptable coherence times and gate fidelities. Each calibration cycle exposes a collection of hardware parameters—such as T₁, T₂, single- and two-qubit gate error rates, and readout errors—that evolve over time due to noise, temperature fluctuations, crosstalk, and hardware ageing.

In practice, these calibration parameters are often:

• monitored one metric at a time,
• inspected using static tolerance bands (e.g., “alert if T₁ < threshold”),
• visualised via dashboards that provide limited supports for multivariate analysis or adaptive behaviour.

This raises a simple question:

Can a lightweight multivariate online anomaly detector provide a more adaptive and informative view of quantum calibration drift than static, metric-wise thresholds?

This repository provides a small, reproducible prototype to explore that question on both synthetic calibration streams and real calibration snapshots obtained via Qiskit.

---

2. Method

2.1 Data: synthetic and real calibration streams

The project treats quantum device calibration as a multivariate time series. Each timestamp is a vector of calibration parameters across multiple qubits.

Two data sources are supported:

• Synthetic calibration data

  • Generated by src/generate_calibration_data.py.
  • Simulates T₁, T₂, gate error, and readout error for a configurable number of qubits.
  • Injects controlled anomalies (e.g., sudden drops in T₁/T₂ or spikes in error rates) to stress-test detectors and support visual inspection.

• Real calibration snapshots (Qiskit)

  • Collected by src/fetch_real_calibration_data.py.
  • Uses qiskit-ibm-runtime to query an IBM Quantum backend for:
    • T₁ and T₂ per qubit,
    • readout error per qubit,
    • single-qubit gate error for a chosen gate (e.g., x).
  • Each execution appends a new timestamped row to data/real_calibration_stream.csv, allowing users to build a real calibration log over time (e.g., by running the script periodically).

Both sources use the same downstream anomaly detection and visualisation pipeline.

---

2.2 Online anomaly detector (Mahalanobis)

The main detector is implemented in src/online_model.py and is designed to be:

• Online/streaming: processes the calibration time series sequentially;
• Multivariate: accounts for correlations across calibration parameters;
• Lightweight: suitable for rapid prototyping and potential integration.

Core components:

1. Sliding window of recent calibration vectors (size configurable).
2. Estimation of mean and covariance on that window.
3. Computation of the squared Mahalanobis distance of the latest observation from the estimated distribution.
4. Maintenance of a short history of Mahalanobis distances and computation of an adaptive threshold as a high percentile (based on a target contamination rate).

A point is flagged as anomalous if its Mahalanobis distance exceeds this adaptive threshold. This yields an online, data-driven boundary between “background” calibration variation and atypical behaviour.

---

2.3 Baseline detectors

To contextualise the behaviour of the online multivariate model, we implement two simple baselines on a selected scalar metric (e.g., T1_q0):

1. Static ±k·σ threshold

   • Given an empirical mean μ and standard deviation σ of the metric, a point is flagged if it lies outside [μ − kσ, μ + kσ].
   • The parameter k can be adjusted in the UI.

2. |z|-score threshold

   • The metric is normalised to a z-score z = (x − μ) / σ.
   • A point is flagged if |z| exceeds a given threshold z_thresh.

These baselines represent typical “rule-of-thumb” monitoring strategies and serve as comparison points for the online Mahalanobis detector in terms of:

• how often alerts are raised (false alarm characteristics),
• whether certain anomalies are detected earlier or exclusively by one approach.

---

3. Demo

The main entry point for interactive exploration is the Streamlit app src/app.py, which provides an end-to-end demonstration of the anomaly detection pipeline.

Running the demo

Install dependencies

git clone https://github.com/freelansire/Quantum-Calib-Anomaly.git
cd Quantum-Calib-Anomaly

pip install -r requirements.txt