
Researchers at IQM Quantum Computers, alongside a collaborator at QbitSoft, have published an empirical benchmarking study evaluating projected quantum kernels (PQKs) across 15, 30, and 45 qubits on the 54-qubit IQM Emerald and 20-qubit IQM Garnet superconducting processors. Detailed in a technical preprint, the team implemented a cross-device referee-challenger framework designed to prevent quantum machine learning (QML) models from achieving inflated accuracy by fitting device-specific hardware noise.
| IQM Projected Quantum Kernel Hardware & Error-Mitigation Benchmarks | ||
|---|---|---|
| Experiment Configuration | Hardware Topology & Circuit Depth | Quantum Challenger vs. Classical Baselines |
| 15 Qubits (Shallow) | Cross-chip (Emerald referee, Garnet challenger; 3 Trotter steps, depth 45) | 81% quantum accuracy vs. 51% best classical model (+30-point statistically significant gap). |
| 15 Qubits (Deep, Unmitigated) | Cross-chip (10 Trotter steps, depth 129, 441 CZ gates) | Quantum accuracy dropped to 68% vs. 64% best classical (+4-point gap; not statistically significant). |
| 15 Qubits (Deep + Error Mitigation) | Cross-chip (10 Trotter steps; REM, ZNE, and Adaptive Pauli Twirling) | REM added 0 points; ZNE restored +14 points via label shifts; APT restored +17 points and halved feature noise. |
| 30 & 45 Qubits | Disjoint qubit chains on IQM Emerald (3 to 10 Trotter steps) | Maintained +17 to +28-point gap over 11 classical ML models; 45-qubit MPS simulation matched QPU (72% vs. 73%). |
Building on the engineered classification protocol introduced by Google researchers, the workflow encodes Fashion-MNIST image components into a second-order Trotterized 1D Heisenberg Hamiltonian. One processor acts as a referee to assign labels maximizing geometric separation against classical radial basis function kernels, while an independent processor or spatially separated qubit chain acts as the challenger, competing against 11 classical baselines including kernel support vector machines, gradient boosting, random forests, and neural networks.
Comparing mitigated hardware runs against noiseless statevector references revealed a critical distinction in error-mitigation behavior: while zero-noise extrapolation (ZNE) widened the measured classification gap by altering the referee’s labeling bandwidth, IQM’s adaptive Pauli twirling (APT) reduced effective feature residuals directly, bringing Emerald’s accuracy on ideal labels to within 2.5 percentage points of a noise-free processor. The authors emphasize that because a classical matrix product state (MPS) simulator with bond dimension 128 reproduced the 45-qubit circuit in under 20 seconds, future experiments targeting genuine quantum advantage will transition to harder-to-simulate 2D quantum dynamics.
Read IQM’s technical blog post here, access the full research preprint on arXiv here, and view the foundational studies by Huang et al. here and Bowles, Ahmed, and Schuld here.
November 17, 2026