Near-Real-Time Global Thermospheric Density Variations Unveiled by Starlink Ephemeris

Abstract

Previous efforts to retrieve thermospheric density using satellite payloads have been limited to a small number of satellites equipped with GNSS (Global Navigation Satellite System) receivers and accelerometers. These satellites are confined to a few orbital planes, and analysis can only be conducted after the data are processed and updated, resulting in sparse and delayed thermospheric density datasets. In recent years, the Starlink constellation, developed and deployed by SpaceX, has emerged as the world’s largest low Earth orbit (LEO) satellite constellation, with over 6000 satellites in operations as of October 2024. Through the strategic use of multiple orbital shells featuring various inclinations and altitudes, Starlink ensures continuous near-global coverage. Due to extensive coverage and frequent maneuvers, SpaceX has publicly released predicted ephemeris data for all Starlink satellites since May 2021, with updates approximately every 8 h. With the ephemeris data of Starlink satellites, we first apply a maneuver detection algorithm based on mean orbital elements to analyze their maneuvering behavior. The results indicate that Starlink satellites exhibit more frequent maneuvers during thermospheric disturbances. Then, we calculate the mechanical energy loss caused by non-conservative forces (primarily atmospheric drag) through precise dynamical models. The results demonstrate that, despite certain limitations in Starlink ephemeris data, the calculated mechanical energy loss still effectively captures thermospheric density variations during both quiet and disturbed geomagnetic periods. This finding is supported by comparisons with Swarm-B data, revealing that SpaceX incorporates the latest space environment conditions into its orbit extrapolation models during each ephemeris update. With a maximum lag of only 8 h, this approach enables near-real-time monitoring of thermospheric density variations using Starlink ephemeris.

1. Introduction

The thermosphere is a layer of Earth’s atmosphere situated between the mesosphere and the exosphere, extending from approximately 80 km to 600 km (or higher) in altitude. This region is a critical zone for human space activities. Despite the extremely low density (on the order of ${10}^{-12} \mathrm{k}\mathrm{g}/{\mathrm{m}}^3$), the residual gas molecules within the thermosphere exert drag on high-speed satellites [1,2]. This drag, known as atmospheric drag, acts in opposition to the satellite’s direction of motion. As a result, atmospheric drag causes mechanical energy loss in satellites, leading to gradual orbital decay, particularly for those in LEO. Variations in thermospheric density, frequently disturbed by space weather activities, significantly impact spacecraft operations, including orbit determination and prediction, space debris monitoring and mitigation.

Thermospheric density exhibits complex spatiotemporal variations, including diurnal, multi-day, seasonal, and long-term changes, driven primarily by solar radiation, geomagnetic activity, and Earth’s rotation and revolution [3,4,5,6,7,8]. Modeling and predicting the thermospheric variations are challenging, requiring extensive observational data. Currently, thermospheric density retrieval primarily relies on two approaches: (1) calculating orbital decay based on orbit tracking data and (2) measuring non-conservative forces using onboard satellite instruments.

The first approach, based on orbit tracking data, utilizes two-line element (TLE) sets provided by the North American Aerospace Defense Command. TLE data cover a wide range of space objects and have long-term tracking records, making them valuable for studying long-term thermospheric variations and atmospheric modeling [9,10,11,12]. However, the irregular and low update frequency of TLE data, coupled with frequent data gaps during geomagnetic storms, limit their ability to capture detailed thermospheric density variations.

The second approach, based on satellite payload data, can be further divided into two categories: (1) direct measurement of non-conservative forces using high-precision accelerometers and (2) indirect calculation of non-conservative forces through precise orbit determination (POD) data obtained from GNSS receivers [13]. These efforts have focused on a limited number of satellites equipped with GNSS receivers or accelerometers, such as CHAMP [14], GRACE [15], GOCE [16], Swarm [17], and GRACE-FO [18]. High-precision accelerometers directly measure non-conservative forces acting on a satellite’s center of mass during orbital motion, enabling the detection of subtle variations of atmospheric drag. Numerous studies have utilized accelerometer-based thermospheric density observations to analyze the responses of thermospheric neutral density to extreme geomagnetic storms and investigate the fine structure of the thermosphere [19,20,21,22,23,24]. However, accelerometers are expensive and prone to technical failures, limiting their widespread use. As orbit determination accuracy has improved, POD data from GNSS receivers have increasingly become a valuable complement to accelerometer-based observations in thermospheric density retrieval. By combining POD data with high-fidelity dynamical models, researchers can estimate orbital decay and non-conservative forces. This approach enables the derivation of thermospheric density with resolution and accuracy comparable to accelerometer measurements [25,26,27,28,29,30]. Despite the above advancements, the limited orbital coverage of these satellites restricts continuous global observations, while delayed updates of POD data (typically several days) and accelerometer data (often a few months) hinder the real-time monitoring of thermospheric density variations.

With advancements in space technology and the expansion of the space industry, satellite launches are no longer limited to scientific research. Numerous companies have proposed mega LEO communication constellations [31], with SpaceX’s Starlink emerging as the largest LEO constellation. SpaceX began launching Starlink satellites in 2019, and as of October 2024, the constellation comprises over 6000 mass-produced small satellites in LEO. This Starlink constellation enables continuous global coverage, but the sheer size of the constellation poses significant collision risks [32,33]. Several near-miss incidents between Starlink satellites and other spacecraft have been reported [34]. In May 2021, recognizing the need for international space safety and the collision risks associated with Starlink satellites, SpaceX began publishing predicted ephemeris data for Starlink satellites on space-track.org approximately every 8 h, facilitating collision avoidance efforts for space activities.

In this study, we conduct a comprehensive assessment of the capability of Starlink predicted ephemeris data to capture thermospheric density variations. First, based on the unique characteristics of Starlink ephemeris data, we first apply a maneuver detection algorithm based on mean orbital elements, which allows us to investigate Starlink’s routine orbital maneuvers and its responses to disturbed thermosphere during geomagnetic storms. Subsequently, based on energy conservation principles, we compute 1-min resolution mechanical energy loss for each Starlink satellite. By analyzing the relationship between mechanical energy loss derived from Starlink ephemeris data and thermospheric density changes from other satellites during both quiet and disturbed geomagnetic conditions, we evaluated the accuracy and reliability of this dataset.

2. Data and Methods

In this study, Starlink ephemerides published by SpaceX on space-track.org (https://www.space-track.org/#publicFiles, accessed on 13 January 2025) were used to calculate the mechanical energy loss of each Starlink satellite. These data feature 72-h predictions and are updated three times a day (even during geomagnetic disturbed periods), with a time resolution of 1 min. Each ephemeris file consists of a descriptive header followed by thousands of ephemeris points spanning the following three days. As detailed in the official Starlink specifications (https://www.space-track.org/documents/SFS_Handbook_For_Operators_V1.7.pdf, accessed on 13 January 2025), these ephemerides conform to the Modified ITC ephemeris format.

Table 1. Starlink ephemeris data example.

Form	Line Number	Content
Header (Line 1–4)	1	created:2024-10-01 10:18:30 UTC
	2	ephemeris_start:2024-10-01 09:33:42 UTC ephemeris_stop:2024-10-04 09:33:42 UTC step_size:60
	3	ephemeris_source:blend
	4	UVW
First Ephemeris Point (Line 5–8)	5	2024275093342.000
		4741.3443204153      2356.2722509008    4125.4543221036
		0.1367078084      6.6298975741      −3.9309088221
	6	7.1787884970e−07   −4.3946653063e−07   8.6816961396e−07
		4.6892685935e−10   −2.8619644577e−10   1.2167821657e−06
		1.2224254104e−09
	7	−1.0345626097e−09   1.7751967007e−12   2.6256571511e−12
		−5.4598689236e−10   4.3500917993e−10   −1.7690294263e−12
		−9.6760401482e−13
	8	5.5649992482e−13    1.3345625691e−12   2.9188509553e−12
		1.6219733961e−09    1.6559550374e−15   −2.8281793013e−15
		5.4168888107e−12

presents the contents of the header and the first ephemeris point from STARLINK-32392 in the batch of 2024-10-01UTC13_21_09_01 as an example. The four-line header specifies the creation time and start and stop times of the ephemeris, along with its source and the covariance reference frame. Specifically, in this example, the header in the table reveals a 60-s interval between consecutive ephemeris points and a total timespan of three days, resulting in 4321 ephemeris points per file. The covariance reference frame is UVW, also known as RTN (Radial, Transverse, and Normal), a spacecraft-centered reference frame used to describe relative motion in orbit. For individual ephemeris points, the first data line contains the epoch timestamp (formatted as “yyyyDOYhhmmss.sss”, provided in Coordinated Universal Time, UTC) and the 3-dimensional position and velocity vectors (in units of km and km/s, respectively, provided in the EME2000 reference frame). The subsequent three lines systematically arrange the corresponding covariance matrix elements.

The currently deployed first-generation Starlink constellation is organized into five orbital shells [35]. Based on orbital altitude and inclination, as shown in

Table 2. Orbital shells of the first generation Starlink.

Orbital Shells	Altitude (km)	Inclination (°)	Plane Counts	Version
Shell 1	550	53	72	v0.9/v1.0
Shell 2	570	70	36	v1.5
Shell 3	560	97.6	6	v1.5
Shell 4	540	53.2	72	v1.5
Shell 5	560	43	28	v1.5/v2.0 mini

, we can categorize each Starlink satellite into its respective shell. After excluding satellites with significant orbital fluctuations and anomalous behavior, the number of operational Starlink satellites (during October 2024) in Shells 1–5 is 1115, 388, 232, 1473, and 1335, respectively.

Taking Shell 1 as an example, Figure 1a illustrates the geographic distribution of satellites within this shell. The satellites are densely concentrated between 53°S and 53°N, consistent with the designed orbital inclination of 53°. In orbital mechanics, the right ascension of the ascending node (RAAN) specifies the orientation of a Keplerian orbital plane in space, while the argument of latitude defines the position of a spacecraft along the orbit. This implies that, over an orbital period, the RAAN remains nearly constant, whereas the argument of latitude varies continuously from 0° to 360°. This characteristic allows us to divide satellites of each shell into distinct orbital planes. Here, we select values of RAAN on 00:00 UT (universal time) 3 October 2024 as a reference to number the orbital planes. For Shell 1, we divide the shell into 72 orbital planes at 5° intervals, as depicted in Figure 1b, where each plane is color-coded to indicate its corresponding region. Note that Starlink satellites in lower-numbered orbital planes cross the equator earlier in local time (LT) than those in higher-numbered orbital planes. Figure 1c presents a histogram quantifying the number of satellites per orbital plane. The results show that all orbital planes are populated, with satellite counts ranging from 8 to 22 and an average of approximately 15 satellites per plane.

If a satellite is influenced solely by Earth’s gravity, its motion around the Earth could be simplified as a two-body problem, and its orbit could be described by constant Keplerian orbital elements. However, in reality, satellites are subject to additional perturbing forces, causing the orbital elements to change gradually. These time-varying orbital elements are referred to as osculating orbital elements. The perturbations in osculating orbital elements can be analytically decomposed into secular, long-period, and short-period variations [1]. Among these, short-period variations are induced by the periodic motion of the satellite along its orbit. In this study, we employ a 20 × 20 Earth gravitational field model to estimate short-period terms [36,37]. Then, the mean orbital elements can be obtained by filtering short-period variations from osculating elements. Compared with osculating elements, the mean orbital elements (especially the mean semi-major axis, SMA) can accurately reflect orbital motions caused by atmospheric drag and orbital maneuvers.

Figure 2a,b display the osculating and mean SMA of STARLINK-1009 from the reference batch, respectively. The osculating SMA in Figure 2a exhibits sinusoidal periodic fluctuations with peak-to-peak amplitudes of up to 15 km, though the changes between peaks are not pronounced. In contrast, the mean SMA in Figure 2b reveals three distinct orbital phases: Natural Decay, Maneuver, and Ambiguity. The Natural Decay phase (blue data points) represents natural orbital decay, characterized by adjacent SMA differences of less than 5 m and a clear downward trend. Conversely, the Maneuver phase (red data points) shows adjacent SMA differences exceeding 10 m, with the mean SMA gradually increasing. The irregular patterns observed in gray data points in the third day align with Liu’s findings, which attribute this phenomenon to SpaceX’s simplified orbital design strategy that considers only ${\mathrm{J}}_2$ perturbations [38]. Additionally, the mean SMA (pink data points) derived from the subsequent batch is also shown in Figure 2b. The decay rates of the Natural Decay phase differ from those of the reference batch, suggesting that each batch is generated from a new orbital extrapolation. Considering the accumulation of errors over time in extrapolation [38] and the updated features of the Starlink ephemeris, we employ a data integration approach in which subsequent batches replace overlapping arcs from previous data. This iterative process continues until a continuous ephemeris sequence is established, as illustrated in Figure 2c,d, where color-coded data represent different batches. Notably, the updated features of the Starlink ephemeris induce discontinuities at batch boundaries that may compromise maneuver detection accuracy. Therefore, we exclude data within transition zones from subsequent analyses.

A satellite’s orbital motion is governed by both conservative and non-conservative forces. From an energy conservation perspective, when a satellite is subjected exclusively to conservative forces—such as Earth gravitational field, tidal perturbations, gravitational influences from the Sun and Moon, and relativistic effects—its total mechanical energy remains constant. In contrast, when non-conservative forces come into play, including atmospheric drag, solar radiation pressure, and Earth’s albedo pressure, the satellite experiences a gradual decrease in mechanical energy. This energy loss corresponds precisely to the work done by these non-conservative forces. Generally, for low-orbit satellites, atmospheric drag accounts for a major proportion of the total non-conservative force in most cases [1,2].

Consider two consecutive epochs, the change in mechanical energy can be expressed as follows [28]:

$$∆E=\left(V_{K_2}-V_{Su{n}_2}-{ V}_{Moo{n}_2}\right)-\left(V_{K_1}-{ V}_{Su{n}_1}-V_{Moo{n}_1}\right)-{\int }_{r_1}^{r_2}{\overrightarrow{g}}_{ECI}·d\overrightarrow{r}$$

(1)

In this equation, the subscripts 1 and 2 denote the parameters corresponding to the preceding and subsequent epochs, respectively. Here, $\overrightarrow{r}$ represents the position of the satellite, and ${\overrightarrow{g}}_{ECI}$ is the Earth gravitational force in the Earth-centered inertial frame. The term $V_K$ is the satellite’s kinetic energy, which can be calculated directly by the velocity $v$ provided in the ephemeris through the equation $V_K={\displaystyle \frac{1}{2}}{v}^2$. $V_{Sun}$ and $V_{Moon}$ corresponded to the central gravitational potentials of the Sun and Moon [39]:

$$V_i={GM}_i\times \left({\displaystyle \frac{1}{s}}-{\displaystyle \frac{1}{\left|{\overrightarrow{r}}_i\right|}}-{\displaystyle \frac{\overrightarrow{r}·{\overrightarrow{r}}_i}{{\left|{\overrightarrow{r}}_i\right|}^3}}\right)$$

(2)

where ${GM}_i$ denotes the gravitational constant of the celestial body (Sun or Moon), $s$ represents the distance between the satellite and the celestial body, and ${\overrightarrow{r}}_i$ represents the geocentric position vectors of the celestial body. The positions of the Sun and Moon can be interpolated from the planetary and lunar ephemerides (DE440) [40] provided by the Jet Propulsion Laboratory.

The last integral term on the right-hand side of Equation (1) represents the cumulative work performed by Earth’s gravitational force on the satellite throughout its orbital motion. The Earth’s gravitational force is computed through the geopotential spherical harmonic model [2] after gravitational potential coefficient corrections for Earth tide effects, including both solid Earth tides and ocean tides. The precise force model incorporates the following components: 120-order gravity field model from EGM2008 [41], the solid Earth tide model from IERS2010 [42], and the ocean tide model from EOT11a [43].

Between consecutive epochs, the satellite trajectory is approximated as a short arc. To ensure the accuracy of the gravitational work integration, an eight-point piecewise Lagrange interpolation scheme is employed to calculate both the Earth gravitational force ${\overrightarrow{g}}_{ECI}$ and the satellite’s geocentric position vectors $\overrightarrow{r}$. Then, the mechanical energy loss ($∆E$) can be derived from the Starlink ephemeris data using Runge–Kutta methods. Notably, the mechanical energy loss of Starlink satellites at 500–600 km will be small during brief observational epochs; thus, errors introduced from imperfect force modeling approximations and cumulative numerical integration inaccuracies are magnified [44]. Nevertheless, our analysis focuses primarily on relative variations in mechanical energy loss. Through strict adherence to the consistent mechanical modeling throughout all computations, the method in this paper can mitigate the effect of the systematic errors while maintaining the computational efficiency required for near-real-time processing of massive satellite constellation data.

It should be noted that Starlink satellites periodically perform orbital maneuvers using their onboard electric propulsion systems, which significantly increase mechanical energy and thus affect the accuracy of mechanical energy loss calculations. To address this, we implement the aforementioned maneuver detection method to exclude data points associated with maneuvers.

After calculating the mechanical energy loss ($∆E$) caused by atmospheric drag, a 9-point moving window median filter is applied to suppress anomalous data points and enhance the statistical robustness. Then, the mean thermospheric density between two epochs ($∆t$) can be computed using the following equation [29]:

$$\rho ={\displaystyle \frac{∆E}{v∆t}}/\left({\displaystyle \frac{C_{d}A}{2m}}{v}_r^2\right)$$

(3)

where $C_d$ is the drag coefficient, $A$ is the effective cross-sectional area, $m$ is the satellite mass, and $v_r$ represents velocity of relative motion between satellite and atmosphere. The accurate calculation of $C_{d}A$ requires detailed knowledge of satellite attitude information and panel design specifications, including surface area and material properties. However, these parameters of Starlink satellites remain unavailable due to SpaceX’s non-disclosure. Therefore, we utilize the mechanical energy loss of Starlink satellites as a proxy indicator for thermospheric density variations.

In this study, we utilize Swarm-B’s thermospheric density data computed by the European Space Agency (ESA), which can be obtained through ESA’s data portal (https://swarmhandbook.earth.esa.int/catalogue/SW_DNSxPOD_2, accessed on 13 April 2025), to validate the results of the mechanical energy loss derived from Starlink ephemerides. We adopt an inverse approach by converting the Swarm-derived thermospheric density into equivalent mechanical energy loss ($∆E$) through parameter substitution in Equation (3). The term $C_{d}A$ is estimated using the Sentman model [45] combined with Swarm satellite specifications documented in ESA’s technical report (https://earth.esa.int/eogateway/documents/20142/37627/swarm-thermo-optical-properties-and-external-geometry.pdf, accessed on 13 April 2025). This methodology enables direct comparison between the Swarm-derived results and those calculated from Starlink ephemeris data.

3. Results

In this section, we analyze the mechanical energy loss results derived from Starlink ephemerides in October 2024. This period comprised two consecutive geomagnetic storms. The first was relatively moderate, whereas the second constituted a strong magnetic storm. Figure 3 presents the interplanetary and geophysical parameters during the storm events. The solar flux proxy F10.7 remained elevated, signifying heightened solar activity. The onset of the first storm was marked by abrupt enhancements in solar wind velocity and density at approximately 06:00 UT on 6 October. Concurrently, the interplanetary magnetic field (IMF) $B_z$ and $B_y$ components exhibited pronounced fluctuations. Subsequently, the AE and Kp indices surged to peak values of 1500 nT and 7, respectively, while the disturbance storm time (Dst) index plummeted to a minimum of −150 nT.

As the effects of the first storm subsided, a strong magnetic storm commenced at around 13:00 UT on 10 October. A Coronal Mass Ejection (CME) shockwave arrived at nearly 12:00 UT, characterized by rapid increases in solar wind velocity (from 658 km/s to 749 km/s within one hour) and density (from 7.1 cm⁻³ to 34.5 cm⁻³). The IMF $B_z$ component displayed extreme southward orientation, reaching a peak value of −42.2 nT at 23:00 UT on 10 October, while $B_y$ oscillated between −22.9 nT and 20.2 nT. Simultaneously, the AE and Kp indices escalated sharply to 3500 nT and 8.7, respectively, accompanied by a positive sudden impulse in the Dst index, identified as a Sudden Storm Commencement (SSC). The Dst index ultimately reached a minimum of −333 nT, signifying extreme geomagnetic disturbance.

As mentioned in the Introduction, satellites in orbit are continuously subjected to atmospheric drag, leading to orbital decay. To maintain their orbital altitude, satellites typically perform regular maneuvers. During geomagnetic storms, thermospheric expansion causes a sharp increase in atmospheric density, accelerating orbital decay and potentially necessitating additional maneuvers. Using Shell 1 data as an example, Figure 4 presents the maneuver detection statistics for Starlink satellites in Shell 1. The results for each orbital plane are shown in Figure 4a, where lower-numbered planes are positioned higher in the figure to maintain feature continuity. The total number of maneuvering satellites across all orbital planes in Shell 1 is shown in Figure 4b. Figure 4a reveals distinct dark blue lines of high maneuver counts spanning all planes, indicating that Starlink satellites follow a predefined maneuver strategy, sequentially maneuvering plane by plane. We refer to these as routine maneuvers. In contrast, irregularly distributed maneuvers (with no clear pattern in timing or orbital plane) are classified as special maneuvers, which we speculate are performed by SpaceX to adjust specific satellite orbital parameters or execute emergency collision avoidance. The statistical results show that, under normal conditions, routine maneuvers begin with satellites in Plane 37 and end with those in Plane 36, occurring approximately every two days.

Between two-day routine maneuvers, similar patterns are also observed. For instance, prior to 8 October, Planes 1–36 executed large-scale maneuvers, while, on 9 and 11 October, maneuvers covering all planes occurred shortly after the previous routine maneuvers. Correlating these events with Dst index variations, we can associate them with geomagnetic storms. Notably, the number of special maneuvers during the period from 8–14 October significantly exceeded that during the quiet Dst period before 8 October. The additional spikes in Figure 4b on 8 and 11 October further corroborate the increased maneuver frequency during thermospheric disturbances. Therefore, the maneuver frequency of Starlink satellites can, to some extent, reflect the level of thermospheric disturbance.

Next, we analyze the thermospheric information embedded in the Starlink ephemeris from the perspective of energy variations. Figure 5 presents the mechanical energy loss derived from Starlink ephemerides and Swarm-B thermospheric density data during 1–30 October 2024. During this period, Swarm-B maintained an altitude of approximately 512 km, with its ascending and descending nodes crossing the equator at LTs of ~09:40 and ~21:40, respectively. The thermospheric density data along the ascending orbit of Swarm-B, computed by ESA, are shown in Figure 5a, while the corresponding Dst index is plotted in Figure 5f for reference. Figure 5a reveals that the thermospheric density during quiet conditions was approximately $3\times {10}^{-12} \mathrm{k}\mathrm{g}/{\mathrm{m}}^3$. During the onset of the first storm (as the Dst index gradually declined), the thermospheric density exhibited an enhancement. In the recovery phase, the density not only recovered rapidly but also dropped below the pre-storm levels. This behavior aligns with the characteristic post-storm overcooling of the thermosphere, attributed to the prolonged recovery time of NO cooling relative to thermospheric temperature and density. As a result, NO cooling continues to operate even after a storm, further cooling the thermosphere [46]. This phenomenon was more pronounced during the subsequent strong storm. At 01:00 UT on 11 October, the Dst index reached its minimum of −333 nT, marking the end of the main phase. While the Dst index recovered to quiet time levels, the thermospheric density remained significantly lower than the pre-storm values until 26 October, persisting for an extended duration.

The mechanical energy loss of Swarm-B, computed from thermospheric density data using Equation (3), is presented in Figure 5b. A comparison with Figure 5a reveals that the variations in mechanical energy loss are nearly proportional to the thermospheric density, with discrepancies primarily attributable to the $C_{d}A$ term, which varies with Swarm-B’s attitude and thermospheric temperature.

This proportional relationship between mechanical energy loss and thermospheric density should also hold for Starlink satellites. The mechanical energy loss derived from Starlink ephemerides is illustrated in Figure 5c,d. Here, we present continuous results for two Starlink satellites: STARLINK-44934 (Figure 5c) and STARLINK-46047 (Figure 5d). Unlike Swarm satellites, which are near-polar orbiters with slowly varying equator-crossing LTs, Starlink satellites, designed with an inclination of 53°, exhibit rapid orbital LT variations. Specifically, the LT of STARLINK-44934’s ascending node at the equator shifted from ~09:40 on 2 October to ~23:55 on 29 October, while STARLINK-46047’s LT transitioned from ~19:40 on 2 October to ~09:40 on 29 October.

Clearly, both Figure 5c,d exhibit a trend of high-value regions shifting from the Southern Hemisphere to the Northern Hemisphere over this month. This phenomenon arises because the region of maximum density along an orbit corresponds to the LT of noon, and as UT progresses, the geographic location associated with a specific LT gradually migrates from the Southern to the Northern Hemisphere. Consequently, when STARLINK-46047 crosses the equator at 12:00 LT, its high-value regions are nearly symmetrically distributed between the two hemispheres, as the results showed on nearly 23 October.

As previously noted, STARLINK-44934 operates primarily at the dawn and nightside (ascending nodes), while STARLINK-46047 is active at the dusk and dayside (ascending nodes). Figure 5c,d reveal that the mechanical energy loss of the former is consistently lower than that of the latter, which aligns with the expected lower thermospheric density at the nightside due to the absence of solar radiation.

Both results of the dayside and nightside indicate responses to the 11 October geomagnetic disturbance, though the dayside manifests a more pronounced reaction (characterized by significant enhancements in mechanical energy loss) compared to the nightside. This day–night asymmetry in thermospheric density enhancement aligns well with the findings of Bruinsma et al. [47].

However, unlike Swarm-B, STARLINK-46047’s response to storms lags by a few hours. This delay stems from the Starlink ephemerides update cycle, which occurs every ~8 h. During each update, the latest space environment conditions are incorporated for orbit extrapolation. Consequently, if a space weather event occurs between two consecutive batches, the resulting data discontinuity manifests as sharp boundaries in the STARLINK-46047 data on 8 October and 11 October, coinciding with the two geomagnetic storms. These results also indicate that, during rapidly evolving storms, discrepancies may emerge between the predicted and actual atmospheric density variations until the next scheduled update.

The post-storm overcooling phenomenon is also evident in the STARLINK-46047 data following the 11 October storm. However, unlike Swarm-B, the thermosphere density reflected in STARLINK-46047 recovered earlier. This difference is attributed to STARLINK-46047’s LT gradually shifting toward noon, resulting in greater solar radiation exposure compared to Swarm-B, which maintained a relatively constant LT of ~09:40.

To further validate the Starlink satellite results, we aligned the LTs with Swarm-B by selecting data from 10 Starlink satellites whose equator-crossing LTs of ascending nodes were approximately 09:40 at different periods. These data were combined to create a continuous dataset, as shown in Figure 5e, where results from individual Starlink satellites are demarcated by red dashed vertical lines.

In Figure 5e, after accounting for LT variations, the mechanical energy loss exhibits patterns remarkably similar to those of Swarm-B in Figure 5b, including comparable responses to geomagnetic storms, recovery rates, and the duration and intensity of the post-storm overcooling phenomenon. This compelling agreement strongly validates the utility of Starlink ephemeris data for thermospheric and space weather monitoring.

4. Discussion

From the above results, the mechanical energy losses derived from individual Starlink satellites demonstrate great consistency with the thermospheric density variations. More significantly, the true strength of Starlink lies in its dense LEO constellation, which provides comprehensive coverage across global and different altitudes, as illustrated in Figure 1a and Table 2. Based on this perspective, in this section, we shift our analysis to the constellation level, and discuss the distinct advantage of utilizing Starlink ephemeris data from the entire satellite network.

Using data from Shell 1 of the Starlink constellation as an example, we selected three distinct periods—before, during, and after the geomagnetic storm—to project the mechanical energy loss of all satellites on a global scale at identical UTs. Contour maps were generated separately for ascending and descending orbits, as shown in Figure 6. Note that the data coverage is constrained to 53°S–53°N due to Shell 1’s orbital design parameters. The corresponding Dst indices for the three periods are marked by vertical lines of different colors in Figure 6g.

First, comparing contour maps of ascending and descending orbits at any period reveal similar patterns of high and low values at the dayside and nightside. However, at finer spatial scales, the mechanical energy loss values for the same geographic location do not fully align. For instance, in Figure 6a,b, corresponding to UT 12:00, 1 October, the high-value region near the prime meridian (driven by solar radiation) exhibits distinct edge patterns: the ascending orbit shows a southwest-to-northeast arc, while the descending orbit displays a northwest-to-southeast arc. Note that the edge of the high-value region is aligned with orbit trajectories. This discrepancy results in a twofold difference in mechanical energy loss at 60°W, 53°S between the two orbits. In reality, the mechanical energy loss experienced by a satellite passing through the same region should be nearly identical, regardless of whether it is in ascending or descending orbit. We attribute this artifact to the usage of predicted ephemeris data rather than post-processed POD data. Since each data point is derived from the preceding one, the results for ascending and descending segments exhibit continuity. This finding is further supported by the closer agreement between the southwestern data of the ascending segment and the southeastern data of the descending segment, as well as between the northwestern data of the descending segment and the northeastern data of the ascending segment, as these regions are temporally contiguous but artificially divided into separate orbital segments.

Second, examining the data for the same orbital segment across the three periods reveals variations in the mechanical energy loss. For any given region, the mechanical energy loss during the storm exceeds that before the storm, which, in turn, exceeds that after the storm. The first inequality is expected, as the energy injection from the storm heats and expands the thermosphere, increasing its density and thereby enhancing the atmospheric drag. The second inequality, however, stems from the aforementioned post-storm overcooling phenomenon.

Next, we compare the mechanical energy loss across different shells of the first-generation Starlink constellation, with the results presented in Figure 7. As before, data from each shell are divided into ascending and descending nodes. Figure 7 displays data from Shell 1, Shell 2, Shell 4, and Shell 5 from top to bottom. Shell 3 is excluded due to its significantly fewer available orbital planes compared to the other shells. In Shell 2, white bars in the colormaps indicate orbital planes that remain unpopulated. For each orbital plane in all the shells, we use the median mechanical energy loss of all satellites within the plane at each timestamp to represent the plane’s data, thereby reducing errors and mitigating the impact of outliers. It is important to note that this approach reduces the dimensionality of the data by eliminating latitude variations, but it facilitates analysis from a constellation-wide perspective.

Overall, the variation patterns of mechanical energy loss are similar across all shells. High-value regions gradually shift from lower-numbered to higher-numbered orbital planes over time, forming an upward-sloping band. This band corresponds to orbital planes where the ascending node crosses the equator around local noon. The apparent inconsistency between ascending and descending nodes within each shell arises from the 12-h LT difference between ascending and descending equator crossings, despite both belonging to the same orbital plane (determined by the RAAN and argument of latitude). This introduces a 180° phase difference. For example, in Shell 1, orbital plane 12 in the ascending nodes shares the same LT as orbital plane 48 in the descending nodes.

Notably, all shells exhibit clear responses to the geomagnetic storm and the post-storm overcooling phenomenon, demonstrating that overcooling is observable across multiple altitudes. Further comparison reveals the following hierarchy of mechanical energy loss: Shell 1 > Shell 4 > Shell 5 > Shell 2. However, the orbital altitude ordering is Shell 4 < Shell 1 < Shell 5 < Shell 2. Generally, higher orbital altitudes result in lower mechanical energy loss due to reduced atmospheric drag, which explains the relationships among Shell 1, Shell 2, and Shell 5. However, Shell 1, despite its higher altitude, exhibits greater mechanical energy loss than Shell 4. We attribute this anomaly to differences in satellite versions between the two shells, as listed in Table 2. Shell 1, deployed much earlier, consists of Starlink v1.0 satellites, while the subsequent shells primarily use the updated Starlink v1.5 design. The v1.0 satellites are comparatively bulkier, potentially experiencing greater atmospheric drag. Additionally, the v1.0 satellites feature visors to reduce albedo by blocking sunlight reflection, while visors were removed for the v1.5 satellites. This design evolution implies uneven solar radiation pressure effects between the two versions, further contributing to the observed anomaly in mechanical energy loss.

The above results generally reflect the variation in thermospheric density at the global scale. Should precise orbital data from Starlink satellites become accessible, they would substantially enhance the spatiotemporal resolution and multi-altitude coverage of thermospheric density measurements. This advancement could enable more refined atmospheric characterization and significantly improve our understanding of thermospheric dynamics.

5. Conclusions

In this study, we conducted a comprehensive assessment of the capability of SpaceX’s publicly available Starlink predicted ephemeris data to capture thermospheric density variations. Based on the unique characteristics of Starlink ephemeris data, a maneuver detection algorithm utilizing mean orbital elements is applied. Statistical analysis of the maneuvers in Shell 1 reveals that satellites perform routine maneuvers covering all orbital planes approximately every two days. In addition to the routine maneuvers, the maneuver frequency of Starlink satellites increases significantly during periods of thermospheric disturbance. This suggests that the maneuver frequency of Starlink satellites can serve as a preliminary indicator of thermospheric disturbance conditions.

The thermospheric density retrieval method based on energy conservation principles is employed to calculate the 1-min resolution mechanical energy loss for each Starlink satellite, and the mechanical energy loss derived from Starlink ephemeris data has been compared with thermospheric density changes observed by the Swarm-B satellite during both quiet and disturbed geomagnetic conditions. The results demonstrate that Starlink’s mechanical energy loss effectively captures thermospheric density variations, although it does not match the accuracy of the results from post-processed POD data.

While Starlink ephemeris data provide valuable insights into thermospheric density variations, they still exhibit certain limitations in accurately capturing these changes. These limitations may arise from factors such as the effective resolution of the dataset. For instance, discrepancies between ascending and descending orbit segments highlight the limitations inherent in the orbit extrapolation model used in prediction ephemeris, which suggests that the extrapolation model is better suited for capturing broad thermospheric trends rather than fine-scale variations.

The Starlink ephemeris represents a vast and valuable database. With its 8-h update cycle, as well as the global coverage and multi-altitude deployment, the Starlink ephemeris enables near-real-time monitoring of the global thermospheric density. Future research can leverage the unique advantages of Starlink ephemeris to advance our understanding of the space environment dynamics and improve the space weather forecasting capabilities.