Day‐To‐Day Variability of the Neutral Wind Dynamo Observed by ICON: First Results From Conjugate Observations

First results are presented from the conjugate maneuvers performed by NASA's Ionospheric Connection Explorer (ICON) spacecraft. During each several‐minute maneuver, ICON crosses the magnetic equator, measuring the plasma drift at the ∼600‐km apex of a magnetic field line and the neutral wind profiles (∼90–300 km altitude) along both ends of that field line. The analysis utilizes 149 pairs of maneuvers separated by ∼24 hr but at nearly the same location and local time. Principal component regression reveals that 39 ± 7% and 24 ± 9% of the day‐to‐day variance in the daytime vertical and zonal drift, respectively, is attributable to conjugate neutral winds. The remaining variance is likely driven by external potentials from non‐conjugate winds and geomagnetic activity (median Kp 2−). Zonal winds at 100–113 km and >120 km altitude are the primary drivers of conjugate vertical and zonal drift variance, respectively. These observations can test vertical‐coupling mechanisms in whole‐atmosphere models.


Introduction
Accurate predictions of day-to-day changes in the ionosphere remain elusive, partially due to influences from the lower atmosphere (Fang et al., 2018;Forbes et al., 2000;Liu, 2016;Liu et al., 2013;Rishbeth & Mendillo, 2001).Various mechanisms transmit lower atmospheric variability to the ionosphere, such as the neutral wind dynamo, field-aligned drag, and composition.The importance of these mechanisms is not well constrained observationally.
In this study, we present observational results quantifying the importance of the neutral wind dynamo, the process by which the kinetic energy of the neutral wind is transformed into electromagnetic energy (Heelis, 2004;Richmond & Thayer, 2000;Rishbeth, 1997;Vasyliunas, 2012).On time scales longer than about a minute, a divergence in the field-line-integrated wind-driven current due to a gradient in the neutral wind or conductivity will lead to a polarization electric field E such that currents are continuous: where the current density j has two terms, one associated with the electric field E and one associated with the neutral wind u, both in the Earth-fixed frame: where σ is the conductivity tensor, and B is the magnetic flux density.Through the dynamo mechanism, a perturbation in the neutral wind can cause a perturbation in the E × B plasma drift.The vertical component of daytime E × B drift at low latitudes is of particular importance because it modulates F-region plasma density by raising or lowering plasma and altering the production-loss-transport balance.
In this paper we distinguish between conjugate and non-conjugate wind dynamo driving.Conjugate driving refers to electric fields developing on a field line in response to winds on the same field line.This is readily quantifiable by NASA's Ionospheric Connection Explorer (ICON), as discussed below.However, nonconjugate winds can also be important (Fesen et al., 2000;Liu, 2020;Liu & Richmond, 2013;Millward et al., 2001;Zhou et al., 2020).It is of interest to quantify how much of the variability in E × B drifts is attributable to conjugate winds.
While numerous modeling studies have investigated wind and electric field variability (see reviews by Liu, 2016;Yamazaki & Maute, 2017;Heelis & Maute, 2020), comprehensive evaluations of these models require coordinated observations of winds and drifts, which are rare (Hysell et al., 2014;Pfaff et al., 2020).The Pfaff et al. (2020) study utilized two daytime sounding rockets, finding that the σ ⋅ E current largely canceled the σ ⋅ (u × B) current, implying a dominance of conjugate dynamo action.
To fill this observational gap, ICON was launched in October 2019, offering the first simultaneous, coincident measurements of electric fields and neutral winds on a global scale (Immel et al., 2018).Immel et al. (2021) presented initial results from the ICON mission, finding correlation coefficients of 0.47-0.56between: (a) changes in the noon-time E × B velocity at ~600 km and (b) changes in a dynamo driving term derived from simultaneous neutral wind observations at the north footpoint of the same field line.These results implied that about 20%-30% of the measured drift variance could be attributed to conjugate wind driving in the northern hemisphere.The unexplained drift variance could be caused by conductivity variability, external forcing from magnetosphere-ionosphere coupling, non-conjugate winds, or of particular relevance to this study, winds at the southern footpoint.Other studies have addressed dynamo driving using ICON (see Immel et al. (2023) and references therein), but they only used its nominal north-looking mode.
ICON was equipped with the ability to occasionally perform a special maneuver to observe the winds at both footpoints of the magnetic field line on which the drift is observed (Immel et al., 2018).Examination of data from this "conjugate maneuver" provides interhemispheric context for the nominal north-looking mode.In this paper, we present first results from this data set to quantify wind forcing from both hemispheres for the first time (i.e., the total conjugate dynamo effect).Specifically, we focus on temporal changes over 24 hr in order to quantify day-today variability.

ICON Data and Conjugate Maneuver Operations
Two ICON data products are used in this study: the Michelson Interferometer for Global High-resolution Thermospheric Imaging (MIGHTI) neutral wind profiles (v5) and the Ion Velocity Meter (IVM) plasma drifts (v6), ICON data products 2.1 and 2.7, respectively.Plasma drifts were measured at the spacecraft (~600 km altitude).Neutral winds were observed remotely at altitudes of 90-300 km using the Doppler shift of the red (630.0nm) and green (557.7 nm) airglow emissions.MIGHTI utilized two identical sensors, MIGHTI-A and MIGHTI-B, to measure orthogonal wind components.The data product descriptions, algorithms, caveats, uncertainties, and validation efforts are described in the embedded documentation and in relevant publications (Englert et al., 2017(Englert et al., , 2023;;Harding et al., 2017Harding et al., , 2021;;Heelis et al., 2017Heelis et al., , 2022;;Makela et al., 2021;Wu et al., 2023).Only IVM drifts with the highest quality are used (RPA_Flag = DM_Flag = 0).
ICON's conjugate maneuver comprises three segments, shown schematically in Figures 1a and 1b).Prior to the maneuver, ICON is in nominal operating mode with MIGHTI-A looking magnetically northeast and MIGHTI-B looking magnetically northwest.In the first stage of the maneuver, ICON yaws 90°(clockwise from above) such that MIGHTI-A and MIGHTI-B observe the line-of-sight (LoS) wind at the southern and northern footpoint, respectively, of a magnetic field line.About 4 minutes later, ICON returns to its normal orientation and IVM samples the ion velocity at the apex of that field line.About four minutes after that, ICON yaws again (90°counterclockwise) to achieve the opposite orientation, such that MIGHTI-A and MIGHTI-B observe the northern and southern footpoints, respectively.Throughout the maneuver, the pitch and roll are controlled to maintain MIGHTI limb-viewing, such that the bottom of the field of view samples a constant altitude (~90 km).
At each position, ICON dwells for 3 MIGHTI exposures, about 90 s, implying 650 km of horizontal averaging due to spacecraft motion.In this analysis, we average these three exposures, because (a) it reduces the effects of photon shot noise and (b) the variation among these three exposures is less than the maneuver-to-maneuver variation, so minimal information is lost.IVM data are also averaged over the 90-s dwell, which ensures that MIGHTI and IVM are probing the same spatiotemporal scales.
At the magnetic equator, the geometry of the ICON observation is ideal for magnetic conjugacy.The tangent location of wind measurements naturally tracks the magnetic field line, for a dipole magnetic field and observation angles of 45°azimuth.In practice, the average absolute miss distance from perfect conjugacy is 4.5°in longitude, which is less than spacecraft motion during 90 s and commensurate with the daytime MIGHTI spatial resolution of several degrees (see Appendix of Harding et al. (2021)).
Conjugate maneuvers were performed whenever allowed by operational and geometrical limitations.For example, power constraints demanded that maneuvers occur with adequate solar illumination; thus, this study pertains to the daytime dynamo.Furthermore, the proper conjugate geometry for MIGHTI could only be attained when ICON flew nearly parallel to the magnetic equator.Due to ICON's orbital inclination of 27°, this was only possible in regions of high magnetic declination in the Pacific and Atlantic sectors (on the descending and ascending orbits, respectively).We focus here on the more numerous Pacific sector maneuvers, which avoid the data quality degradation in the South Atlantic Anomaly.

Data Processing
We analyze 23 Nov 2020-14 Nov 2022, which includes all conjugate maneuvers excluding data before Oct 2020 and during morning local times (LT < 11 hr) because of potential IVM photoemission contamination and/or tenuous plasma, as discussed by Heelis et al. (2022).Of the 151 available maneuvers, 3 were rejected by a 3σ outlier filter used to detect spurious drifts.
A coordinate conversion is performed on the LoS winds at each altitude to calculate zonal and meridional wind, using the same procedure implemented for ICON data product 2.2 during normal science mode.Red-line (greenline) observations are used above (below) 170 km.Data above 290 km, close to the top edge of the MIGHTI field of view, are less accurate and are removed from the analysis.This leaves 76 samples in each altitude profile.
Winds and drifts are converted to a common magnetic coordinate system, specifically the modified apex coordinate system (Emmert et al., 2010;Richmond, 1995;van der Meeren et al., 2018).The perpendicular-to-B neutral winds are u 1 and u 2 (corresponding to u e1 and u e2 in Richmond (1995)) which are positive eastward and positive equatorward/downward, respectively.Drift velocities are similarly converted to v 1 and v 2 .Since the "2" direction is nearly vertical at the magnetic apex, we hereafter refer to v 2 as the "vertical drift", noting that it is positive downward.The end results of this processing are the zonal and meridional neutral wind profiles (~90-300 km) along both ends of a field line, and the E × B drift at the apex of that field line, for each conjugate maneuver.Examples are shown in Figures 1c-1l).

Results
In this study, we analyze how the ionospheric velocities, v 1 , v 2 , are related to the wind drivers, u 1 and u 2 , in each hemisphere.Specifically, we investigate If the engineering and orbit geometry constraints allowed a conjugate maneuver to occur, it was often the case that a conjugate maneuver was also allowed ~1 day later, when the spacecraft crossed the equator at nearly the same longitude.Thus, a database is available of pairs of maneuvers separated by 14 (~22.5 hr) or 15 (~24 hr) orbits.
Quantifying the change in the data between these maneuvers provides a measure of day-to-day variability.Additionally, by restricting the analysis to changes over ~24 hr, we mitigate the effects of systematic errors in MIGHTI or IVM data, and we minimize aliasing by seasonal or local-time trends.
The pairs with a 14-orbit lag represent samples at almost the same local time, separated by ~20 deg of longitude and 22.5 hr of clock time.The pairs with a 15-orbit lag represent samples separated by ~0.7 hr of local time and 12 deg of longitude, but almost exactly 24 hr of clock time.Neither set represents the perfect sampling that unambiguously quantifies true day-to-day variability; however, these are the closest approximations possible using orbital sampling.For each of the 148 available maneuvers, a search was conducted to find another maneuver 14 or 15 orbits later (allowing a maneuver to participate in two pairs if possible).In total, 149 pairs of maneuvers were identified.
Figures 2a and 2b show the spatial and temporal sampling used in this study, and 2c shows the Kp index (Matzka, Stolle, et al., 2021) for the entire period and for the specific times the maneuvers occurred.Kp ranged from 0 to 6− , with a median of 2− .Panels d and e show the spatial and temporal lags associated with the pairs to be analyzed.

Correlation Analysis
Figure 2f shows a simple analysis comparing the 1-day change in vertical drift, Δv 2 , with the 1-day change in the zonal wind, Δu 1 , at 106 km altitude at the northern footpoint.Figure 2g shows the same using southern-footpoint winds.The set of 14-orbit and 15-orbit lags are shown separately, but since no significant difference is seen in these results, these sets are combined in subsequent analyses to improve statistics.The Pearson correlation coefficient is +0.51 for north-footpoint winds and +0.48 for south-footpoint winds (p < 0.001, two-tailed t-test).A positive correlation is consistent with physical expectations: an eastward wind in a region with Hall conductivity yields an eastward component of σ ⋅ u × B current, which is balanced by the Pedersen currents associated with a westward electric field (i.e., an equatorward/downward E × B drift).This analysis is repeated at each altitude, and the correlations are shown in Figure 3a.Strong positive correlations are seen in a region around 105 km, negative correlations are seen from ~120-170 km, and positive correlations are seen above.It is noteworthy that the north and south correlation profiles are so similar, suggesting that magnetic-zonal wind control of the vertical drift is hemispherically symmetric.
Figure 3b shows the same but for magnetic-meridional wind.Given the use of magnetic apex coordinates, for which magnetic-meridional is defined as positive downward/equatorward, the differing patterns in the lower thermosphere appear to suggest a surprising asymmetric forcing.However, as shown later, this spurious feature is due to self-correlation.
Figures 3c and 3d show similar analysis for zonal drift.The dominant feature is a positive correlation with eastward winds throughout the middle and upper thermosphere.This is consistent with the classical F-region dynamo: eastward winds cause a vertical polarization of the ionosphere, associated with an eastward E × B drift.
Spurious correlations provide a challenge with interpreting these plots in terms of physical drivers.Generally speaking, zonal and meridional wind changes are correlated with each other and themselves, as a complex function of altitude.In the next section we perform an analysis which accounts for these correlations.

Principal Component Regression
Here, linear regression is used as an alternative to correlation analysis.The coefficients of linear regression are partial correlations, inherently accounting for self-correlations.However, because this data set has fewer samples (149 pairs of maneuvers) than features (304 features per maneuver-a 76-element profile for zonal and meridional winds at each footpoint), linear regression cannot be directly applied.Thus, we apply dimensionality reduction using principal component analysis (PCA) before regression, a procedure known as principal component regression.PCA identifies an orthogonal basis that maximizes the wind variance captured by a finite number of components.The regression analysis will also quantify the amount of drift variance explained by the winds (R 2 ).
The problem is cast as standard linear prediction of vertical drift changes from wind changes: where u is a 304-element vector comprising the four wind profiles (zonal and meridional, north and south), and c is a coefficient vector to be determined.The "Δ" symbols have been omitted for brevity.An analogous equation is constructed for v 1 .PCA is used to optimally approximate u as the sum of a small number (denoted N pc ) of principal components: where the columns of the 304 × N pc matrix A are the principal components, and w is the weight vector.Stacking the four wind profiles into one vector captures zonal-meridional and north-south correlations.In practice, u is quite compressible; for example, ~50% of the wind variance is captured by just the top 5 principal components.Applying this to Equation 3 yields: This form is a standard linear prediction, like above, with a new coefficient vector c T A) T .This new prediction problem is tractable if N pc is small enough.
A choice of N pc that is too large can cause overfitting.To choose N pc , we randomly split the data set into two halves, the training set and the test set.The training set is used to perform the regression and find the coefficients c*, and these coefficients are used to compute the performance (R 2 ) on both sets.If c* represents real physical drivers, then R 2 on the two sets should be comparable.If overfitting is dominating, R 2 will be significantly different.Results are shown as a function of N pc in Figures 4a and 4b.Values and errorbars represent the mean and standard deviation of 1,000 randomized train/test splits.Performance on the test set diverges from the training set and even declines at high N pc , suggesting overfitting.Test-set R 2 is stable for a large range of N pc between about 3 and 15.For the subsequent analysis, we use N pc = 6, a choice which captures the most signal variance while ensuring differences between the training and test sets are statistically insignificant.There are no qualitative or statistically significant differences in any results using N pc from 3 to 10.
We convert the principal component regression coefficients to the altitude domain by multiplying by A. We split this vector into its 4 components for each of the wind terms, and show these 4 coefficient profiles in Figures 4c  and 4d.These coefficients can be thought of as the sensitivity of the vertical drift to the wind, as a function of altitude.
Figures 4c and 4d are similar to the correlation profiles in Figures 3a and 3b, except the zonal wind features above ∼120 km and the entire meridional wind profile are strongly suppressed, suggesting those features were spurious correlations and not physical dynamo drivers.The major physical driver appears to be the zonal wind at 100-113 km (as quantified by the full-width at half-maximum) in both hemispheres.
The profile in Figure 4c traces out the approximate shape of the Hall conductivity profile, consistent with the firstorder approximation of dynamo driving by conductivity-weighted winds.However, it is possible that the shear at ∼115 km suggests a connection with sporadic E, whereby winds could affect the conductivity.Further physical interpretation of these profiles would require detailed modeling.
Figures 4e and 4f show the same analysis applied to zonal drift prediction.The dominant feature, as in Figures 3c  and 3d, is the zonal wind in the middle/upper thermosphere.A small feature is seen at 100 km in the zonal wind and 115 km in the meridional wind, but these are not significant relative to the error.
Generally, these results suggest that zonal winds govern drift variability more than meridional winds.This makes sense physically; although meridional winds are just as variable as zonal winds, their effective magnitude is multiplied by the sin of the inclination angle (∼0.5 for these field lines).Additionally, the meridional wind is only electrodynamically effective insofar as it has an asymmetric component (i.e., a homogeneous northward wind spanning both hemispheres creates no net forcing, but a homogeneous eastward wind creates maximal forcing).Nevertheless, nonzero coefficients are seen in the meridional winds in the lower thermosphere.They are significant relative to the error, but they are of smaller magnitude than the zonal wind coefficients.It is possible that they are an overfitting artifact, but some common features are seen between the hemispheres (especially Figure 4f at 110 km), suggesting a robust result.The much larger data set of North-looking magnetic conjunctions could be used to further investigate the electrodynamic role of meridional winds, though it is expected to be minor relative to zonal winds.

Geophysical Research Letters
10.1029/2023GL107110 The squared coefficients are a quantitative measure of the relative influence of winds at different altitudes on drift variance.Integrating the squared coefficients below and above 150 km-an approximate boundary between the E and F regions (Maute & Richmond, 2017)-shows that 77 ± 7% (55 ± 10%) of conjugate wind drivers are below 150 km for the vertical (zonal) drift.
Figures 4g and 4h show total explained drift variance (R 2 , gray bar) compared to two independent analyses where only the north or only the south footpoint winds are used as inputs to the regression (colored bars).The primary result of this analysis is that 39 ± 7% of the day-to-day vertical drift variance is accounted for by conjugate wind forcing, as determined by test-set performance.The zonal drift R 2 is lower, with only 24 ± 9% accounted for.Predictability is greater when winds in both hemispheres are considered, but an individual hemisphere provides more than half of the relevant wind information.This result is consistent with the Immel et al. ( 2021) analysis of normal-mode ICON data, which found correlations of ∼0.5 between vertical drifts and north-footpoint winds (i.e., R 2 = 25%).Here, the North-only R 2 is 29 ± 9% (Figure 4g, green bar).

Measurement Uncertainty
In general, noise or other short-term uncertainties in the data could artifically reducing the explained variance.This effect is quantified below.
For IVM data, the main error sources are short-term offsets in the drift-meter zero baseline, possibly caused by variations in the spacecraft's electrostatic environment (see Section 2 of Heelis et al. (2022)).This is ameliorated by using data after Oct 2020.Nevertheless, we quantify its effect by computing the 24-hr-averaged drifts on the dates maneuvers occurred and calculating the 1-day change in this "baseline drift," which is a combination of drift-meter offsets and long-period geophysical activity (e.g., planetary waves) (Wu et al., 2023).The standard deviation of this difference across the 149 pairs of maneuvers is 1.9 m/s.Dividing this by the standard deviation of Δv 2 (8.3 m/s) and squaring it, we conclude the amount of vertical drift variance attributable to IVM drift-meter offsets is at most 5.5%, smaller than our statistical uncertainty on R 2 .Zonal drifts are determined primarily by the retarding potential analyzer and are minimally affected by drift-meter offsets.
For MIGHTI winds, the relevant error sources are quantified by the reported "1_sample" and "1_day" precision variables added in quadrature (Englert et al., 2023).We performed an end-to-end Monte Carlo analysis using 1,000 trials with random noise added commensurate with this precision.The reduction in R 2 was 1.9% for vertical drift and 1.0% for zonal drift.

Summary and Conclusions
This study presented initial results from ICON's conjugate maneuvers, the first ever data set of neutral winds spanning the dynamo region in both hemispheres.With magnetically conjugate observations of neutral winds and plasma drifts, it was shown that 39 ± 7% of the day-to-day vertical drift variance and 24 ± 9% of the day-to-day zonal drift variance are attributable to conjugate neutral winds through the daytime dynamo mechanism.Both hemispheres contribute to drift variability, but a single hemisphere provides more than half of the predictive power.
A principal-component-based regression was used to determine the wind drivers responsible for day-to-day drift variability.This showed that zonal winds exert more control than meridional winds, with the vertical drift mostly controlled by the zonal wind at 100-113 km and the zonal drift mostly controlled by the zonal wind above ∼120 km.Notably, this is a purely observational confirmation of theoretical expectations that zonal winds in the Hall-conductivity region and the Pedersen-conductivity region control vertical and zonal drift variability, respectively.
These results imply that 61 ± 7% and 76 ± 9% of the measured day-to-day variance in the vertical and zonal drift, respectively, is not explained by conjugate neutral winds.Observationally accounting for this missing variance is a major goal for future investigations.This unexplained drift variance is likely attributable to the effects of unmeasured winds on field lines with lower and higher apex heights, as well as geomagnetic activity and possibly conductivity variability.In general the influence of conductivity variability is thought to be much smaller than wind variability (Forbes et al., 2019;Maute et al., 2012).Geomagnetic activity was quiescent but non-negligible (median Kp 2 ).The disturbance dynamo effect, insofar as it perturbs conjugate winds, is included in the reported R 2 , but the non-conjugate disturbance dynamo and the penetration electric field contribute to the unexplained variance.Although restricting the data set to low Kp may increase R 2 , the size of the data set does not support such analysis.Future analysis of ICON's conjugate observations could probe other temporal and spatial scales and compare with numerical models.They can also provide quantitative context for ICON's more numerous north-looking observations, such as extending the results shown here to quantify seasonal dependencies.
processed conjugate maneuvers has been made separately available on Zenodo for ease of use (Harding, 2023), comprising the zonal and meridional winds and in-situ ion velocities in modified apex coordinates, for each maneuver used in this study.Kp data were obtained from GFZ Potsdam (Matzka, Bronkalla, et al., 2021).

Figure 1 .
Figure 1.(a) Schematic representation of ICON's conjugate maneuver.Within a span of ∼8 min, observations are made of the ion drift and magnetically conjugate neutral wind profiles (90-300 km).Black dots and dotted lines indicate ICON's location over time.Small arrows indicate the modified apex coordinate system.(b) Topdown view.(c)-(g) Example winds from 5 separate conjugate maneuvers.(h-l) Conjugate drifts from the same maneuvers.
(a) to what extent the change in drift over ~1 day can be attributed to a change in the winds, and (b) what features of the neutral wind control changes in the drifts.

Figure 2 .
Figure 2. (a),(b) The spatial and temporal coverage of conjugate maneuvers.(c) Corresponding Kp index.(d),(e) Spatial and temporal lags of the pairs of analyzed maneuvers.Pairs of maneuvers separated by 14 and 15 orbits quantify changes over ∼1 day but with nearly the same location and local time.(f) Comparison of 1-day changes in the vertical drift, Δv 2 and the zonal wind, Δu 1 at 106 km at the north footpoint, with Pearson correlation.(g) Same for the south footpoint.

Figure 3 .
Figure 3. (a) The altitude dependence of the correlations in Figures 2f and 2g.Correlations are computed between 1-day changes in downward drift at the magnetic apex and magnetic-zonal wind, for both the north and south footpoints.Error bars are standard error of the Pearson correlation.(b) Same as (a) for magnetic-meridional wind.(c),(d) Same as (a),(b) for magnetic-zonal drift.

Figure 4 .
Figure 4. (a) The amount of day-to-day downward drift variance that be explained by conjugate winds, using principal component linear regression applied to conjugate observations.(b) Same for zonal drift.(c), (d) The regression coefficients for downward drift prediction.Coefficients for zonal (c) and meridional (d) wind are plotted separately.(e), (f) Same for zonal drift.(g) The amount of day-to-day downward drift variance explainable using only North or only South footpoint winds, as compared to a case when both are used.(h) Same for zonal drift.Panels c-h use 6 principal components.Values and errorbars are the mean and standard deviation of 1,000 randomized train/test splits.