ml-quantum-error-correction - SKILL.md Agent Skill

name: ml-quantum-error-correction description: > Machine Learning approaches for Quantum Error Correction (QEC). Use when researching, designing, or implementing ML-assisted QEC systems including: (1) diffusion models for error decoding (DiffQEC pattern), (2) reinforcement learning for QEC control and calibration, (3) neural network decoders for surface codes and LDPC codes, (4) loss-biased fault-tolerant architectures, (5) quantum error correction for quantum machine learning (QML). Activation keywords: ML QEC, diffusion model quantum error, RL quantum control, neural decoder, quantum error correction machine learning, QEC decoder, fault-tolerant quantum computing ML, QML model validation, quantum mutation testing, quantum certified training, QNN robustness.

ML-Assisted Quantum Error Correction

Core Patterns

Pattern 1: Diffusion Models for QEC Decoding (DiffQEC)

Use diffusion models to learn the noise-to-error mapping for quantum error correction.

Workflow:

Generate training data from quantum circuit simulations with realistic noise models
Train a diffusion model to map syndrome measurements to error configurations
Use the trained model as a fast decoder during QEC cycles
Validate logical error rate against baseline minimum-weight perfect matching (MWPM)

Key design choices:

Condition diffusion on syndrome patterns (binary vectors from stabilizer measurements)
Use classifier-free guidance to balance decoding speed vs accuracy
Target sub-microsecond decoding for real-time QEC cycles

Pattern 2: Reinforcement Learning for QEC Control

Use RL to adaptively control QEC parameters under environmental drift.

Workflow:

Define state as syndrome history + calibration parameters
Define actions as parameter adjustments (gate durations, pulse shapes, bias ratios)
Reward = negative logical error rate (or proxy via syndrome entropy)
Train with PPO or SAC on simulated quantum hardware

Advantage over recalibration:

Continuous adaptation without halting computation
Learns drift patterns specific to hardware instance
Reduces calibration overhead by 10-100x

Pattern 3: Loss-Biased Fault Tolerance

Exploit physical error bias (loss errors dominant over other errors) for simplified QEC.

Workflow:

Characterize error channel: identify dominant error type (e.g., atom loss in neutral atoms)
Design code that corrects dominant error with fewer resources
Use fast detection mechanism (e.g., autoionization for loss detection < 1ms)
Combine with standard QEC for residual errors

Platforms where this applies:

Neutral atom qubits (Rydberg platforms)
Photonic qubits (loss-dominant channel)
Superconducting qubits with engineered dissipation

Pattern 4: Neural Decoders for Topological Codes

Replace classical decoders (MWPM, union-find) with trained neural networks.

Architecture patterns:

CNN on 2D syndrome lattice for surface codes
Graph neural network for irregular code geometries
Transformer for temporal syndrome sequences

Training data:

Simulate Pauli noise at various physical error rates
Generate syndrome-error pairs for supervised learning
Augment with realistic noise models (crosstalk, leakage, SPAM errors)

QEC for Quantum Machine Learning

When applying QEC to QML workloads:

Error detection vs correction: QML may tolerate higher error rates than universal QC
Mid-circuit measurement: QEC must not destroy quantum state used for ML inference
Code choice: CSS codes preferred for compatibility with variational circuits
Overhead estimation: Factor QEC overhead into quantum advantage calculations

Note: For QML model quality assurance (mutation testing, robustness analysis, certified training via IBP, hardware readiness), see references/qml-model-validation.md. This covers validation of trained QML models, which is distinct from hardware-level error correction.

Evaluation Metrics

Metric	Target	Notes
Logical error rate	< 10^-6	Below threshold for fault tolerance
Decoding latency	< 1 μs	Must be faster than QEC cycle time
Training data size	10^6-10^8 samples	Depends on code distance and noise model
Generalization	Works at unseen p_phys	Must extrapolate beyond training error rates

References

DiffQEC (2604.24640): Diffusion models for versatile QEC
Loss-biased QEC (2604.21876): Fast autoionization for sub-ms QEC cycles
RL control of QEC (2511.08493): Adaptive calibration under drift
QEC for QML (2601.07223): Error correction at the quantum-classical ML intersection
Linear-time QEC codes (2603.04543): Breakthroughs in quantum LDPC decoding