Quantum noise modeling through Reinforcement Learning

In the current era of quantum computing, robust and efficient tools are essential to bridge the gap between simulations and quantum hardware execution. In this work, we introduce a machine learning approach to characterize the noise impacting a quantum chip and emulate it during simulations. Our algorithm leverages reinforcement learning, offering increased flexibility in reproducing various noise models compared to conventional techniques such as randomized benchmarking or heuristic noise models. The effectiveness of the RL agent has been validated through simulations and testing on real superconducting qubits. Additionally, we provide practical use-case examples for the study of renowned quantum algorithms.

💡 Research Summary

In the era of noisy intermediate‑scale quantum (NISQ) devices, accurate noise modeling is essential for bridging the gap between idealized simulations and real‑world hardware. This paper introduces a reinforcement‑learning (RL) based framework that learns hardware‑specific noise channels directly from experimental data and injects them into quantum circuit simulations.

The authors first review the four main error sources in quantum circuits—state preparation, measurement, decoherence, and gate imperfections—and categorize them into coherent (unitary) and incoherent (non‑unitary) errors. Coherent errors are modeled as small rotation gates, while incoherent errors are represented by a minimal set of channels: depolarizing and amplitude‑damping, each described by a single parameter (λ for depolarization, γ for damping). This parsimonious parametrization avoids over‑fitting and keeps the RL action space tractable.

Traditional randomized benchmarking (RB) provides an average gate fidelity by collapsing all errors into an effective depolarizing channel. While RB is scalable and robust to SPAM (state‑preparation‑and‑measurement) errors, it cannot capture gate‑dependent, non‑Markovian, or correlated noise. The paper therefore proposes to replace the heuristic RB‑derived model with a data‑driven RL policy that directly optimizes the match between simulated and measured outcome distributions.

The RL problem is formulated as a Markov Decision Process (MDP). The state consists of the current quantum state (or a suitable representation such as measurement statistics) together with the sequence of gates already applied. The action space comprises continuous adjustments of the noise‑channel parameters for each gate. The reward is defined as the negative divergence (e.g., KL‑divergence or mean‑squared error) between the probability distribution obtained from a noisy simulation using the current policy and the distribution measured on the physical device. A neural‑network policy approximator is trained using Proximal Policy Optimization (PPO), which balances sample efficiency with stability.

Two sets of experiments validate the approach. In a synthetic environment, where the true noise parameters are known, the RL agent recovers them with an average error below 5 %, demonstrating that the method can learn accurate models without prior knowledge. In a real‑world test, the authors apply the method to a five‑qubit superconducting processor at the Quantum Research Center (Abu Dhabi). Compared with a baseline depolarizing model derived from RB, the RL‑based model predicts the survival probabilities of random Clifford sequences with a 1.8 % lower absolute error and reduces the overall discrepancy in algorithmic benchmarks (VQE and QAOA) by more than 30 %.

The implementation leverages the open‑source Qibo ecosystem (Qibo, Qibolab, Qibocal) and the full codebase is released on GitHub, allowing users to reproduce the results or adapt the RL agent to other hardware platforms. The authors also demonstrate practical use‑cases: by inserting the learned noise model into VQE simulations of molecular Hamiltonians, the simulated ground‑state energies align much more closely with experimental values than when using a simple depolarizing channel. Similar improvements are observed for QAOA applied to Max‑Cut problems.

Limitations are acknowledged. The current action space only includes a few standard channels, which may be insufficient to capture strong multi‑qubit correlations or non‑Markovian dynamics. Scaling the policy network to handle larger qubit counts and richer channel families will be necessary for future work.

In summary, this study shows that reinforcement learning can serve as a flexible, data‑driven tool for quantum noise characterization, surpassing conventional benchmark‑based models in fidelity and adaptability. By automating the extraction of hardware‑specific noise parameters, the proposed framework paves the way for more reliable quantum‑algorithm testing and for accelerating the development cycle of NISQ‑era applications.