NoiseVault: calibrated noise from real quantum computers for circuit simulation

Dev D. Goyal

Version 0.3.0, released 4 October 2026. Apache-2.0. Python 3.11 to 3.14.

Abstract

Quantum computers make errors, but most simulators ignore them. NoiseVault adds the measured errors of a real quantum computer to your simulations. One file, a profile, works in Qiskit, Cirq, PennyLane and Stim. A fingerprint pins the exact numbers, so others can rerun your work. The package includes 25 devices from IBM, Quantinuum and Google.

Ideal simulator0 wrong

IBM Fez profile0 wrong

Figure 1.The same three-qubit circuit, run 1,024 times. Each square is one run. Red squares are wrong results.
Figure 2.(a) Each qubit at its place on the chip. Lines connect qubits that interact. (b) The 156 values, on the same color scale. Data: IBM Quantum, 26 February 2025.
Figure 3.(a) The GHZ circuit. (b) Results of 1,024 runs in Qiskit Aer. Dashed outlines show the ideal result.
Table 1.nv check ibm_fez. The dot shows deviation ÷ tolerance. Below 1 is a pass. A pass tests the conversion, not the hardware.
Figure 4.The 256 bits of each fingerprint. Red cells changed.
Table 2.The report of the Qiskit conversion of IBM Fez, shortened.
Figure 5.Simulated counts, 4,000 runs for each circuit. (a) TVD, the distance to the counts (0 is identical). Grey bars: sampling noise. (b) Fitted factors. Triangles: true values.
Figure 6.(a) The median errors of each device. Larger dots have more qubits. (b) The date of each calibration.

Device noise

Every qubit (The basic unit of a quantum computer. A qubit holds 0, 1 or a mix of both.) on a real chip makes errors at its own rate. The manufacturer measures these rates in a calibration (A regular measurement of the error rates of each qubit. The rates change from one calibration to the next.).

On IBM Fez, the readout error (The chance that a measurement gives the wrong bit.) goes from 0.15% to 10.6% across its 156 qubits.

Simulation with a profile

An ideal simulator (A program that calculates the results of a quantum computer.) adds no errors. In a GHZ circuit (A circuit that links qubits so that their measurements always agree.), all three qubits must agree, so each run gives 000 or 111.

With the IBM Fez profile, 27 of 1,024 runs give a wrong result.

One profile, four simulators

Each simulator describes noise in its own format. NoiseVault converts one profile (One file with the measured error rates of one device on one date.) into all four.

nv check confirms that each conversion matches an exact reference calculation.

A fingerprint for each profile

Each profile has a fingerprint (A SHA-256 hash: a 256-bit code calculated from every number in the profile.). Change one number by 1%, and 116 of the 256 bits change.

Put the fingerprint in your paper. If the numbers differ, the load stops with an error.

A report of each conversion

A simulator cannot always represent every effect. Each conversion comes with a report (A list, saved with your results, of what the noise model includes.) of what is exact, approximated or missing.

Comparison with a real device

nv compare scores a profile against counts from a real device. It fits two factors (How much the gate errors and the readout errors must be multiplied to match the counts.), each with a 95% interval.

Here the counts are simulated with known factors, and both intervals contain the true values. No real-device counts have been compared yet.

25 devices included

The package includes 25 devices: 18 from IBM, 5 from Quantinuum and 2 from Google. They load offline, with no account.

nv pull gets newer calibrations from IBM.

Table 3.All bundled profiles. Select a column heading to sort.

Get started

NoiseVault needs Python 3.11 to 3.14.

  1. Try it without an install

    This needs uv. It prints the IBM Fez profile.

  2. Install for your simulator

    Use cirq, pennylane or stim in place of qiskit.

  3. Simulate with real noise

    The circuit from Figure 3. It runs offline.

Limitations

The models use average error rates and treat each error as independent. They do not include:

  • CrosstalkAn operation on one qubit disturbs its neighbors.
  • LeakageA qubit leaves its two working states.
  • Coherent errorsSmall rotation errors add up in a systematic way.
  • Idle noiseErrors while a qubit waits, unless the circuit has explicit delays.
  • Correlated readoutReadout errors that depend on each other.

The model approximates the hardware. It is not a copy. The limitations page gives the full list.

Citation

Cite the software and the fingerprint of each profile that you used.

Data sources

  1. IBM Quantum. Fake backends in qiskit-ibm-runtime. Apache-2.0.
  2. Quantinuum. Hardware specifications repository. Apache-2.0.
  3. Google Quantum AI. Calibrations in cirq-google. Apache-2.0.