Magnitude-dependent Quantum Advantage in Forbush Decrease Detection: A Quantum Kernel SVM Benchmark

Forbush decreases (FDs) — transient reductions in galactic cosmic ray intensity driven by interplanetary coronal mass ejections — are key observables in space weather monitoring, yet their automated detection from multivariate solar wind and neutron monitor time series remains a challenging classification problem. Here we report a systematic benchmark of quantum kernel support vector machines (QKSVM) for FD detection, in which the FD magnitude threshold emerges as the governing factor separating two distinct classification regimes. Using 2971 confirmed events from the Forbush Effects and Interplanetary Disturbances (FEID) catalogue, combined with hourly OMNI solar wind parameters — including interplanetary magnetic field (IMF) components, solar wind speed, proton density, proton temperature, and the Kp and Dst geomagnetic indices — and galactic cosmic ray count rates from the Jungfraujoch neutron monitor station (JUNG, NMDB), we construct a balanced FD versus quiet-time classification dataset and extract 121 statistical features across eleven physical channels. A ZZFeatureMap quantum kernel with 4–8 qubits is benchmarked against a classical radial basis function (RBF) SVM across 180 experimental configurations spanning FD magnitude thresholds of 0%–7%, circuit depths of 1–3 repetitions, and quantum training sizes of 50–250 samples. We find that below a magnitude threshold of 4%, the classical kernel consistently outperforms the quantum alternative (mean ΔAUC=−0.100 at min_magn =0%). Above this threshold, the relationship inverts: at min_magn ≥4% the quantum kernel achieves positive mean ΔAUC in 72% of configurations, rising to 100% of configurations at min_magn ≥5% (mean ΔAUC=+0.074, peak AUC=1.000 with 4–8 qubits), indicating that the entanglement structure of the ZZFeatureMap captures non-linear correlations between IMF dynamics and cosmic ray modulation that the RBF kernel cannot represent. The magnitude threshold of 4% thus constitutes a physically interpretable boundary between a noise-dominated regime where classical methods suffice and a signal-rich regime where quantum kernels provide measurable and statistically significant advantage (p<0.001, Wilcoxon signed-rank test). These results establish FD magnitude as a key predictor of quantum classification performance, and suggest that near-term quantum machine learning applications in heliophysics should preferentially target high-amplitude space weather events.