Predictability of IRI-2020 topside electron temperature errors over the African sector: nonlinear dynamics and out-of-sample skill from Swarm Langmuir-probe observations
Abstract. The residual between the electron temperature measured by the Swarm A and B Langmuir probes and the IRI-2020 topside electron temperature (TBT-2012 option) was analysed along satellite passes over the African low-latitude sector (20° W – 55° E, dip latitude within ± 40°) in a geomagnetically quiet window (February 2024) and a disturbed window (late April to May 2024, including the superstorm of 10–12 May), and predicted out of sample from February to May 2024. Three instrumental effects were identified and handled before the analysis: a relaxation after each 128-s probe sweep, the absence of the radar-based calibration over much of the daytime anomaly, and switches of the night-side temperature between two states separated by 600–900 K that are not reflected in the quality flags; 452 of 627 passes in the two windows were free of such switches. Delay embedding, recurrence quantification and comparison with 100 iterative amplitude-adjusted Fourier-transform surrogates show that the determinism of the residual exceeds that of a matched linear process in 396 of the 452 passes. The excess is largest on the quiet day side (median 5.3 surrogate standard deviations) and smaller under storm forcing (3.3; p = 6.2 × 10⁻⁸), whereas raw determinism increases. Applying the radar-based calibration changes determinism by 0.14–0.32 of the between-pass interquartile range, and replacing TBT-2012 by the Bil-1985 option by 0.15–0.63, so the invariants are attributable mainly to the IRI error in its default configuration. Surrogate-supported along-track horizons of about 330–740 km were found for 51 passes. In the prediction experiment, IRI-2020 alone had an RMSE of 559 K, dominated by an offset that a 60-day climatology of recent passes reduced to 271 K; relative to this climatology, a gradient-boosted tree ensemble using the residual of previous days improved the mean-square error by 0.14 (95 % interval 0.02–0.23), and a nowcast also using the measured ion density by 0.39 (0.27–0.48). Day-side passes with a larger nonlinear excess were predicted less well.