The POES/MetOp satellites (POES hereafter) cover all L shells and several MLT sectors by orbiting with high inclination (∼98.7°) at ∼800–850 km of altitude, with periods of ∼100 min (e.g., Evans and Greer, 2004; Rodger et al., 2010) and providing data at a 2-s cadence. The Medium Energy Proton and Electron Detector (MEPED) onboard each satellite monitors electron and proton fluxes at several energy ranges and two look-directions (0° telescope pointed at zenith and 90° telescope orthogonal to it; 30° of full aperture). With this configuration and a loss cone angle of ∼60° at LEO, when POES crosses mid-to-high latitudes, MEPED allows probing of the outer radiation belt population, both deep into the loss cone (locally precipitating population) and just outside it (locally trapped and mirroring particles) (e.g., Nesse Tyssøy et al., 2016). When intense precipitation is observed in POES/MetOp data, the 0° flux approaches the 90° flux. In other words, given a certain flux of the mirroring population, the portion of the precipitating population is comparable to the trapped one, such that the ratio R = 0°/90° (i.e., precipitation intensity or efficiency) approaches a value of ∼1. When R = 1, precipitation is isotropic, and the loss cone is full. Recent work by Selesnick et al. (2020) demonstrated that the 0° telescope sometimes detects trapped particles when diffusion is weak; however, such ambiguity does not apply to our work as we only consider time intervals of rather intense and distinct precipitation.

The nominal integral electron channels measure electrons at >30 keV (E1), >100 keV (E2), and >300 keV (E3), with the addition of a virtual electron channel that measures electrons at >700 keV (E4) from the P5 (2.5–6.9 MeV) and P6 (>6.9 MeV) proton channels (details in Green, 2013; Yando et al., 2011). Several past studies relied on the combination of these channels or the virtual E4 channel itself to identify relativistic electron precipitation (Capannolo et al., 2019; 2022a; Carson et al., 2012; Chen et al., 2023; Gasque et al., 2021; Qin et al., 2018; Shekhar et al., 2017; 2018; Yahnin et al., 2016; 2017). We also use the differential proton channels onboard MEPED (P1: 30–80 keV, P2: 80–250 keV, P3: 250–800 keV) to investigate the concurrent proton precipitation during the REP events.

It is worth mentioning a few caveats about POES data. In this work, we use the IGRF (International Geomagnetic Reference Field) magnetic field model, which is readily available in POES data. With another more sophisticated magnetic field model, the nightside L shell values would be slightly higher than those reported here. POES is known to have a rather high noise floor level (Nesse Tyssøy et al., 2016) and thus is not that sensitive to low flux values. As a result, our dataset might likely be biased to REP events with moderately high fluxes compared with other REP events observed with more sensitive instruments (e.g., ELFIN, FIREBIRD-II, etc.).

Figure 1 illustrates two examples of a typical REP: wave-driven in panel A and FLCS-driven in panel B. For a wave-driven REP (Figure 1A), the precipitating >700 keV electron flux (red solid line) is enhanced well within the outer belt, marked by the locally trapped >700 keV electron flux (red dashed line). For a FLCS-driven REP (Figure 1B), as L shell increases (from right to left), the first population reaching isotropy (i.e., similar precipitating and trapped flux) is the most energetic one (>700 keV, red); this is then followed by the >300 keV electron IB (green), the >100 keV electron IB (black), and finally the >30 keV electron IB. As a result, the classic signature of a FLCS-driven REP shows high-energy precipitation at lower L than low-energy precipitation, which instead occurs at higher L shells. This is a direct consequence of the electron gyroradius being energy-dependent. High-energy electrons have a larger gyroradius, thus are scattered by field lines with a larger curvature radius (i.e., farther away from Earth), but low-energy electrons, given their smaller gyroradius, require a smaller curvature radius (i.e., closer to Earth) to be scattered.

Capannolo et al. (2022b) developed a classifier of REP events based on the long short-term memory (LSTM; Hochreiter and Schmidhuber, 1997) deep learning architecture. This tool identifies REP and classifies it into either wave-driven or FLCS-driven REP. Although the performance is suitable for identifying and classifying events between wave and FLCS drivers (F1 ∼ 0.95), false positives or misclassified events are still possible. To use this classifier for scientific research, we post-process the model outputs to ensure events are properly classified. The post-processing routine is as follows:

1. Shift by 3 data points for each event (to improve centering the event boundaries around the event and account for the observed LSTM delay; see details in Capannolo et al., 2022b)

Note that none of the identified events occur within the South Atlantic Anomaly. Given that the classifier is based on machine learning, which is intrinsically probabilistic, the event boundaries represent regions of highly likelihood for precipitation, rather than precisely identifying flux enhancements using specific thresholds, as done in previous studies (e.g., Capannolo et al., 2022a; Carson et al., 2012; Gasque et al., 2021). Following the post-processing, we visually inspected each event identified and classified by the model (∼10,000 wave-driven, ∼12,000 FLCS-driven; see Supplementary Table S1 in Supplementary Material) and discarded any non-ideal REP event. An ideal wave-driven event resembles the one shown in Figure 1A, while FLCS-driven events are similar to that in Figure 1B. Specifically, a wave-driven event occurs a) within the outer belt (90° flux is relatively high both at lower and higher L shells than the 0° flux localized enhancement), b) isolated in L shell, and c) without energy-dependent precipitation at E1, E2 or E3. A FLCS-driven REP event is ideal if a) it occurs at the outer edge of the outer belt, b) precipitation is isotropic at all energies within the event boundaries, c) no additional precipitation is occurring during the energy dispersion profile (this could indicate additional waves/mechanisms), and d) no E1, E2, E3 0° flux fluctuations are occurring at L shells higher than the outer event boundary (considering the first ∼5 data points following the event boundary; this ensures the FLCS-driven isotropy is relatively in a steady state). Events categorized as one but belonging to two different classes (waves vs FLCS) are also excluded, but a wave-driven event near an FLCS-driven event (if clearly distinct) is included in the dataset if each event adheres to the aforementioned rules of the respective category. Examples of excluded events are shown in Supplementary Figure S1 in the Supplementary Material. We adopted a system of flags to distinguish between events to keep (flag = 0), discard (flag = 1), events to merge (flag = 2), misclassified events (flag = 3), and events to merge that have been misclassified (flag = 23). Supplementary Table S1 in SM illustrates how many events per flag we found and provides the model performance after our visual filter. The dataset is available in the repository by Capannolo and Staff (2024b).

Table 1 shows the number of wave-driven and FLCS-driven REP events identified, listed by year. There is a much larger number of wave-driven events (∼7,400) than FLCS events (∼2,000), although the model originally identified a similar number of REP in the two categories. We found that most wave-driven events (∼73%) are truly ideal, while FLCS-driven REP tends to be rather complex and does not often adhere to our definition of an “ideal FLCS-driven REP event” (only 16% of FLCS events are included). This finding is not surprising since the tail region is highly dynamic and overlapping mechanisms can be at play (field line scattering, excitation of waves, injections, etc.). We preferred to discard a large number of events in this category, including only those truly driven by FLCS. This approach allows us to study the properties of REP specifically driven by FLCS without the influence of other competing processes. Like any statistical dataset, this one is not necessarily a complete dataset of all REP events occurring from 2012 through 2023, as it relies on the deep learning classifier described in Capannolo et al. (2022b) and adheres to the criteria described above.

We primarily focus on the westward auroral electrojet (AL) index. AL has been widely used to investigate substorm activity and we expect wave-driven or FLCS-driven REP to occur in association with substorms (i.e., during tail stretching and injections). We use the 1-min SML (maximum westward auroral electrojet) and SMR (symmetric ring current intensity) indices, which are the SuperMAG equivalents to the auroral index AL and the high-resolution ring current index Sym-H, respectively (Gjerloev, 2012; Newell and Gjerloev, 2011). We calculate SML* (SMR*) as the minimum SML (SMR) index over 3 h preceding the REP UT. While SML provides an instantaneous measurement of the westward auroral electrojet, SML* is useful to highlight if a substorm was occurring in the 3-h window before the observed REP. Similarly, SMR* provides insights on a storm occurring in the previous 3 h. The OMNI dataset provides 1-min resolution solar wind (SW) data.

The top row of Figure 2 shows the distribution in L-MLT bins (1 L by 1 MLT) of the total number of events (A), the wave-driven events (B), and the FLCS-driven events (C) in a logarithmic color scale. Most events are found in the 18–24 MLT sector, with a peak around ∼21 MLT and primarily focused between 4 and 6 L shells. The panels in the lower row show the occurrence rate of the REP events, calculated as the number of events found in each bin and divided by the number of POES passes in the same bin. The overall trends remain, though these plots highlight that REP events are observed only occasionally by POES data. Considering the total number of REP events and the cumulative days from 2012 through 2023, we find that the POES constellation crosses a region of precipitation at least twice a day (on average).

Wave-driven precipitation occurs at any MLT sector, though is observed more frequently over ∼15–02 MLT, peaking in the heart of the outer belt at 4–6 L shells. This result agrees with previous literature both from POES data as well as other LEO satellites and is often attributed to EMIC wave scattering (e.g., Angelopoulos et al., 2023; Blum et al., 2015; Capannolo et al., 2021; 2023; Gasque et al., 2021). Wave-driven precipitation over 02–14 MLT has also been associated with EMIC waves (e.g., Blum et al., 2024; Hendry et al., 2016; Qin et al., 2018); however, this causal relationship seems to be less strong than that in the post-noon to post-midnight sectors. We do not explore this possible association in this work, though we speculate in section 4 on its simultaneous occurrence with proton precipitation–a proxy of EMIC wave activity. FLCS-driven precipitation only occurs on the night side, where field lines are indeed likely undergoing stretching. The FLCS occurrence rate peaks at pre-midnight (∼21–22 MLT) and at 5–6 L shells, in overall agreement with previous work linking field line curvature scattering with electron precipitation (e.g., Capannolo et al., 2022a; Comess et al., 2013; Smith et al., 2016; Yahnin et al., 2016; 2017; Wilkins et al., 2023). The FLCS-driven occurrence rates are lower than the wave-driven ones given the lower number of purely FLCS-driven events than the wave-driven ones (see Section 2.2 for details), rather than a true indication of FLCS occurring less frequently than wave-driven precipitation. Supplementary Figure S2 in the SM illustrates the distribution of events as a function of latitude and longitude, both in geographic and geomagnetic coordinates.

Figure 3 illustrates the distribution of average electron fluxes and the precipitation intensity (top row for wave-driven and bottom row for FLCS-driven). The average electron flux is calculated by averaging the E4 90° and 0° fluxes for each event and then sorting them into L-MLT bins and calculating the average values. The trapped flux (panels A and D) decreases as a function of L shell and is constant over MLT except for a slight enhancement over 6–10 MLT (Supplementary Figure S3 in SM), reproducing an expected trend for energetic electrons (Qin et al., 2024; Meredith et al., 2016; Allison et al., 2017). The precipitating flux (panels B and E) follows a similar trend in L shell without a clear MLT variation (Supplementary Figure S4 in SM). The fluxes for wave-driven events are overall higher than those during FLCS-driven events. Such a finding is expected as wave-driven REP typically occurs within the outer belt, while FLCS-driven precipitation occurs at the outer boundary of the belt, where the flux is already decreasing.

Panels C and F illustrate the precipitation efficiency or intensity (e.g., Capannolo et al., 2019; Qin et al., 2024), calculated as the ratio of the precipitating flux over the locally trapped flux (fluxes are averaged within the event boundaries) for each event and binned in L-MLT. This ratio estimates how many electrons are precipitating (i.e., deep into the loss cone) compared to those locally mirroring (i.e., outside the loss cone), thus not contributing to the local precipitation. Previous studies also show how this value can be linked to diffusion coefficients, wave properties, and minimum resonant energy (Angelopoulos et al., 2023; Li et al., 2013; Longley et al., 2022); however, these calculations are left as future work. Given the isotropic nature of the FLCS-driven REP, the ratio is high throughout the region where FLCS events are found. Similar to FLCS-driven intensity, wave-driven REP is more efficient as a function of L but presents a minimum over ∼6–12 MLT and ∼3–6 L (Supplementary Figure S5 in SM). The trend in L shell is consistent with previous results from ELFIN observations described by Qin et al. (2024) and is probably due to the steeper L shell slope of the trapped flux compared to the precipitating flux.

An interesting feature of wave-driven REP is that its efficiency drops in the dawn-to-noon MLT sector. Although precipitation in this sector does not occur frequently (Figure 2E), it is nevertheless observed albeit with weaker intensity. This suggests that the dawn-to-noon waves are not particularly efficient at scattering >700 keV electrons. On the contrary, the ratio stays consistently higher elsewhere. The precipitation from noon to post-dusk has often been associated with EMIC waves (e.g., Blum et al., 2015; Capannolo et al., 2021; 2023; Hendry et al., 2016; Li et al., 2014; Rodger et al., 2015): the wave-electron resonant conditions are indeed favorable in these regions of high plasma density and low magnetic field, typically when the minimum resonant energy can be low enough to be detected by the >700 keV POES integral channel (Jordanova et al., 2008; Meredith et al., 2003; Silin et al., 2011; Summers and Thorne, 2003; Qin et al., 2020; Woodger et al., 2018). This would also explain why the efficiency is lower over the dawn-to-noon MLT sector: here, the resonant condition for EMIC-driven precipitation typically occurs at several MeV rather than the preferential sub-MeV and ∼MeV energies detected by POES. In this region, other waves, such as hiss and chorus waves are present, and we cannot exclude their contribution (e.g., Blum et al., 2015; Ma et al., 2021; Reidy et al., 2021; Shumko et al., 2021; Kandar et al., 2023). Identifying the specific wave driver of this precipitation requires further investigations. Precipitation across midnight has also been associated with EMIC waves (Blum et al., 2024; Capannolo et al., 2022a; Comess et al., 2013; Yahnin et al., 2016; 2017; Smith et al., 2016); however, here, both waves and FLCS contribute to precipitation, with the FLCS-driven efficiency being higher than the wave-driven one.

The DL-based classifier (mentioned in Section 2.2) identifies the boundaries of each REP event, typically characterized by intense precipitation. Here, we calculate the radial extent ΔL and ΔMLAT (magnetic latitude) associated with each event. ΔL estimates the approximate equatorial region in the radial direction where waves or FLCS are efficient at scattering electrons, while ΔMLAT provides the latitudinal extent at low altitudes. To avoid bias in the analysis, we also rule out a small percentage of events that span only a single data point ( ≲ 9% wave-driven and ≲ 1% FLCS-driven). Visually, these events are more extended than only one data point. Overall, we noticed that the DL classifier tends to be conservative in estimating the extent of the events, and thus the precipitation scales might be slightly underestimated. It is also worth noticing that the boundaries of REP are typically somewhat arbitrary as they can depend on the precipitating flux or the precipitation efficiency (different studies give different definitions to infer the radial scales). Figure 4 illustrates the ΔL (left) and ΔMLAT (right) properties. The top panels (A–D) indicate the radial extents binned in L and MLT (bins of 1 L and 1 MLT widths) and the lower panels (E, F) show the histograms. Radial scales are overall localized (<0.3 L, <1° MLAT), in agreement with previous studies (e.g., Capannolo et al., 2021; 2023; Gasque et al., 2021; Woodger et al., 2018). Wave-driven REP is more localized (average: 0.16 ΔL, 0.53° ΔMLAT; median: 0.13 ΔL, 0.41° ΔMLAT, standard deviation: 0.13 ΔL, 0.44° ΔMLAT) than FLCS-driven REP (average: 0.18 ΔL, 0.53° ΔMLAT; median: 0.17 ΔL, 0.51° ΔMLAT; standard deviation: 0.08 ΔL, 0.21° ΔMLAT), with a longer tail at higher radial scales. Across midnight, where FLCS and waves are both contributing to the precipitation, the FLCS-driven REP is broader than the wave-driven REP in both ΔL and ΔMLAT.

There is a clear asymmetry between dayside and nightside for wave-driven REP, with REP being broader on the dayside than the nightside, as evident in both radial and latitudinal scales (see Supplementary Figures S6, S7 in SM for more details). Again, this could be an indicator that waves or the scattering regions are more extended on the dayside, possibly a consequence of different generation mechanisms (magnetotail injections vs solar wind fluctuations). Several case studies leveraged multi-point observations and found that dayside EMIC waves triggered by solar wind structures could be more extended in both MLT and L shell (e.g., Blum et al., 2016; 2021; Engebretson et al., 2015; 2018; Usanova et al., 2008; Yan et al., 2023a; 2023b; Yu et al., 2017), while nightside waves are generally more localized, often occurring during substorm activity (e.g., Blum et al., 2015; Capannolo et al., 2019; Clilverd et al., 2015; Jun et al., 2019a; Jun et al., 2019b). This was also statistically confirmed by Blum et al. (2017) through measurements by Van Allen Probes. They found that dayside EMIC waves are more spatially extended than nightside EMIC waves, which instead tend to persist longer. Furthermore, Figures 4A, C reveal a minor asymmetry pre/post-midnight for wave-driven REP. As previously found in Capannolo et al. (2022a), post-midnight REP is more localized than pre-midnight REP, possibly suggesting that the waves or the conditions favorable for electron scattering vary in radial scale across midnight. Contrary to the day/night asymmetry, the variation across midnight has yet to be explained.

Finally, we want to emphasize that we only consider the spatial scale of single REP events as identified by a single POES pass across the precipitation region. There are several indications that REP occurs in patches, covering multiple MLT sectors, likely reproducing the L-MLT extent of its associate driver. For example, previous case studies show several satellite passes or balloon observations associated with EMIC wave activity, spread over a few MLT sectors (e.g., Capannolo et al., 2021; Shekhar et al., 2020; Woodger et al., 2018), demonstrating that the entire region of REP is certainly broader than that observed by a single POES pass. Similarly, when the magnetotail stretches away from Earth, we expect that a few MLT sectors will be affected by FLCS, likely delineating a nightside REP that extends in longitude (Wilkins et al., 2023; Zou et al., 2024; Sivadas et al., 2019). Accurately quantifying the realistic extent over MLT (not only in the radial/latitudinal direction) is a key step in estimating the true energy input into the atmosphere, which we aim to explore in future studies. In the next section, however, we present a first approximation.

As discussed in the introduction, REP can impact the atmospheric chemistry and possibly the radiative balance. Its effects heavily depend on the energy input into the atmospheric system, defined not only by the energy flux but also the precipitation spatial extent (in latitude and longitude), as well as its duration. As a first comparison, Figure 5 illustrates the contribution to the atmosphere due to waves (top) or FLCS (bottom). The first column shows the fraction of precipitating flux depending on the associated driver compared to the total precipitating flux. Waves dominate the precipitation over the FLCS, contributing to at least 70% of the average precipitating flux in regions overlapping with FLCS-driven REP.

Out of the factors that quantify the REP energy input (flux, size, duration), we can estimate the >700 keV input power (a combination of energy flux and spatial extent) assuming a) the 0° electron flux is constant throughout the loss cone, b) the center energy for the >700 keV channel is ∼879 keV (Peck et al., 2015), c) the latitudinal extent of REP is calculated as the difference between the minimum and maximum magnetic latitude in each bin, and d) the longitudinal extent is assumed ∼1 MLT (bin size of the dial plots). The power is calculated as the >700 keV precipitating electron flux multiplied by the center energy, the solid angle factor for a loss cone of ∼58° (2π[cos (0°)-cos (58°)] ∼ 2.96 sr) and the spherical area covered by the latitude (as in c) above) and longitude (1 MLT) extent of the REP (in each bin). The results are in Figures 5B, E, highlighting that the input power for wave-driven REP is systematically higher than FLCS-driven REP, mostly due to the higher energy flux during wave-driven precipitation. Panels C and F illustrate the input power weighted by the occurrence rate of REP in Figures 2E, F. Energy deposition most often occurs at L > 5 in the pre-midnight region for FLCS-driven REP and from post-noon to midnight for wave-driven REP (peaking over 5–7 L), with a smaller contribution at 9–11 MLT.

Assuming an EMIC-driven precipitation region of 1° in magnetic latitude and 3–12 MLT azimuthal extent (Blum et al., 2017; 2020; Clausen et al., 2011; Engebretson et al., 2015; Hendry et al., 2020; Kim et al., 2016a; Mann et al., 2014), Capannolo et al. (2024a) estimated, from a small sample of EMIC-driven precipitation events observed by ELFIN (Capannolo et al., 2023), an average hemispheric contribution of a few to 10s of MW, with an energy flux in the loss cone of ∼3.3 × 10⁻² erg/cm²/s (63 keV–2.8 MeV electron energies), primarily deposited in the mesosphere. Wilkins et al. (2023) estimated an average energy flux varying from ∼0.1–0.6 × 10⁻² erg/cm²/s and a contribution of ∼10 MW for FLCS-driven >50 keV precipitation (area defined by 1° latitudinal extent and 18–06 MLT) using ELFIN. For the present dataset, the average energy flux in the loss cone is ∼1.33 × 10⁻² erg/cm²/s for a wave-driven REP and ∼0.4 × 10⁻² erg/cm²/s for a FLCS-driven REP, with an average latitudinal extent of ∼0.5° (calculated as point c) above), providing an average input power of 0.66 MW and 0.19 MW (considering 1 MLT of azimuthal scale), respectively. For wave-driven REP, assuming an azimuthal extent of 3–12 MLT, the input power is ∼2–8 MW. For FLCS-driven REP, assuming an azimuthal extent of 2–10 MLT (the highest boundary given the distribution of events in Figure 2, third column), the power is ∼0.4–2 MW. These estimates are comparable to those from previous results (Capannolo et al., 2024a; Wilkins et al., 2023) albeit smaller given a more localized radial extent and POES higher orbit (∼847 km on average) compared to ELFIN’s (∼450 km), resulting in a smaller loss cone (∼58° vs ∼66°) thus energy flux. Furthermore, these estimates only include the electron flux >700 keV and, due to the high noise level affecting POES, the electron fluxes above a few MeV are likely underestimated. Finally, the 0° telescope only probes deep into the loss cone over a field of view of 30°, underestimating the total loss cone flux.

From a power standpoint, wave-driven REP is clearly dumping more energy into the atmospheric system; however, providing only the input power is not yet enough to quantify the total energy input as atmospheric effects of REP significantly depend on the duration of such phenomenon. Once more light is shed on how sustained wave-driven vs FLCS-driven REP is, one can finalize the entire energy input into the atmosphere and perform modeling to quantify the associated effects (e.g., Duderstadt et al., 2021). In addition, while wave-driven REP mostly occurs at >700 keV possibly accompanied by 100 s keV electrons if driven by EMIC waves (Capannolo et al., 2021; 2023; Hendry et al., 2017), FLCS drives precipitation across all energies, down to 10 s keV, thus affecting a broader range of altitudes, from the E region and below.