betaFeedback

WorksPhysics and Astronomy

The Neural Density Model: An Open 3D Thermospheric Density Reanalysis for 2003–2026 From Tracked Space Debris

Abstract

Thermospheric mass density controls satellite drag but is observed directly by few missions, and the leading operational product (HASDM) relies on non-public radar data. We present the Neural Density Model (NDM), a 3D density reanalysis for October 2003–October 2026 at 250–800~km. Previous approaches either convert the orbital decay of tracked objects into point densities or correct empirical models with a few global scale factors. NDM instead fits a spatially resolved neural correction to NRLMSISE-00 directly to the observed decay of 600–800 objects per year, through a differentiable drag operator, jointly with their ballistic coefficients, calibration spheres and accelerometer densities. On withheld satellites the median error is 0.078 (natural-log units), against 0.101 for HASDM and 0.310 for NRLMSISE-00; NDM has the smaller bias, HASDM slightly smaller scatter. On Spire GNSS orbits, never used in training, NDM and HASDM perform comparably. NDM uses only public observations.

main.pdf

Key Points:

  • We present NDM, an open 3D thermospheric density reanalysis for 2003 to 2026 at 250 to 800 km, built only from public observations

  • A differentiable drag operator fits a neural density correction directly to the orbital decay of 600 to 800 tracked objects per year

  • On withheld satellites the error is 23 percent below HASDM and fourfold below NRLMSISE-00; on independent Spire orbits it matches HASDM

Corresponding author: Enrico Camporeale, [email protected]

Plain Language Summary

Satellites in low Earth orbit are slowed by the thin upper atmosphere, and how strongly depends on its density, which can double within hours during geomagnetic storms. Predicting satellite positions, avoiding collisions and planning re-entries all depend on knowing this density, but only a handful of research satellites have ever measured it directly. We turn the space debris and spent rocket stages already tracked by radar into a global network of density sensors: as they lose altitude, the rate of their decay reveals the density they have flown through. A machine-learning model, trained jointly on the decay of 600–800 such objects per year and on the few direct measurements, reconstructs the density between 250 and 800 km every day from 2003 to 2026. Tested against satellites it never saw, it is about as accurate as the best operational model, which is built on restricted military data, and four times more accurate than the standard empirical model. This open record can serve as a reference for satellite operators and as a training set for machine-learning forecasts of space weather.

Introduction

Atmospheric drag is the largest and least predictable non-gravitational force on objects in low Earth orbit (LEO), and its uncertainty is dominated by the uncertainty in thermospheric mass density [12, 3]. With more than ten thousand active satellites and a growing debris population, density errors propagate directly into conjunction assessment, re-entry prediction and constellation operations. The February 2022 loss of 38 Starlink satellites during a moderate storm [16, 2] and the large drag-induced orbit changes during the May 2024 superstorm [30] illustrate the stakes.

Empirical models such as NRLMSISE-00 [33] and NRLMSIS 2.1 [13] have systematic errors of tens of percent, reaching a factor of about two at solar minimum, and their storm-time errors are larger still [6, 29, 12]. Direct observations are sparse: accelerometer and precise-orbit densities from CHAMP, GRACE, GOCE, Swarm and GRACE-FO [11, 39, 44, 38] sample only a few orbital planes and altitudes at any time. Data-assimilation studies have corrected empirical or physics-based models using accelerometer densities [18, 17, 35, 7, 9], GNSS orbits of CubeSats [41], or drag of a few objects [5, 28], and machine learning has been used to emulate, calibrate or forecast density [27, 25, 20, 45]. Most of these efforts adjust a handful of global drivers or scale factors of an existing model over days to a year. The spatial pattern of the density, that is, how it varies with latitude, local time and altitude, therefore remains that of the underlying model. Reduced-dynamic and energy-dissipation retrievals from single satellites [36, 8] are accurate along their own orbits but are not global.

The orbital decay of the objects routinely tracked by the US Space Surveillance Network, distributed as two-line element sets (TLEs), is a free, continuous and global record of drag [32, 15, 10]. The operational state of the art, the High Accuracy Satellite Drag Model (HASDM; [40]), also uses the observed orbital decay of tracked objects to correct an empirical model. It dynamically calibrates a Jacchia-type empirical model (originally Jacchia 1970, JB2008 in current versions; [22, 4]) using radar observations of several tens of calibration objects. HASDM densities are available to researchers on request ([42]), but the radar observations, the calibration-object set and the processing are not public, so the product cannot be independently reproduced, extended or audited.

Weather forecasting was transformed by reanalyses such as ERA5 ([19]): long, gap-free, observation-constrained reconstructions of the atmospheric state that serve both as a climatological reference and as training data for machine-learning forecast models. No equivalent open product exists for the thermosphere. Here we present the Neural Density Model (NDM), a 23-year 3D density reanalysis built entirely from public or openly licensed data. Three elements make it possible: (1) a neural parameterization of the density field whose parameters are fitted by gradient descent through a fully differentiable drag observation operator, so that the orbital decay of each object constrains the density field directly, without first being converted into point estimates of density; (2) training on the orbital decay of 600–800 rocket bodies and debris objects per year (about 2.8×1062.8 \times10^{6} drag intervals over 2003–2026), anchored in absolute scale by spheres of known area-to-mass ratio and by accelerometer densities; and (3) a validation design that tests the product only on data it has not seen, including the GNSS orbits of the Spire constellation, which never enter training, and that separates the contribution of each data source.

Method

Neural Parameterization of the Density Field

NDM represents the log-density as a correction to NRLMSISE-00:

ln⁡ρ(x,t)=ln⁡ρMSIS(x,t)+δ(x,t),δ=cd+fθ(x,zd),(1)\ln\rho(\mathbf{x},t) = \ln\rho_{\mathrm{MSIS}}(\mathbf{x},t) + \delta(\mathbf{x},t), \qquad\delta= c_d + f_{\theta}(\mathbf{x},z_d), \tag*{(1)}

where dd is the day containing tt. The input vector x\mathbf{x} contains latitude, local solar time and day of year (as sines and cosines), altitude, daily and 81-day F10.7, KpK_p, and the current and lagged 3-hourly apa_p (previous 9 h, 12–33 h and 36–57 h). The network is a multilayer perceptron (3 hidden layers of 128 units, SiLU activations). The daily level cdc_d and a four-dimensional latent vector zdz_d are free parameters fitted to that day’s observations; they carry the day-to-day state that the indices do not explain, while fθf_\theta encodes how that state maps onto latitude, local time and altitude. To keep the daily state from jumping between days, the loss includes a random-walk penalty

R=λs∑d[(cd+1−cd)2+∥zd+1−zd∥2],(2)\mathcal{R} = \lambda_s \sum_d \left[(c_{d+1}-c_d)^2 + \lVert z_{d+1}-z_d \rVert^2\right], \tag*{(2)}

which penalizes day-to-day changes unless the data require them; days with few observations then borrow information from their neighbors. The weight λs\lambda_s is selected for each year from {1,10,100}\{1, 10, 100\} on validation weeks. NRLMSISE-00 itself is evaluated with its full geomagnetic input, the 3-hourly apa_p history of the preceding 57 h (its storm-time option), rather than with the daily ApA_p alone. One network is trained per calendar year (with ±15\pm15-day overlaps) as an ensemble of three seeds; the yearly fields are concatenated into the reanalysis.

The observations therefore constrain the field on time scales of one day and longer, through (cd,zd)(c_d,z_d) and the 1–60-day TLE intervals. Variability on shorter time scales enters only through the geomagnetic indices in x\mathbf{x} and through NRLMSISE-00: it is model-driven, with a response learned from the observations, not observed directly.

A Differentiable Drag Observation Operator

Each TLE gives the mean motion nn of an object, the average angular rate of its orbit (for example in radians per second), which is related to the semi-major axis aa by n2a3=μn^{2}a^{3}=\mu, where μ\mu is Earth’s gravitational parameter. Drag removes orbital energy, so the orbit shrinks and nn increases. For an object ii with consecutive TLEs at times t1t_{1} and t2t_{2}, the observation is therefore the change in mean motion, Δniobs=n(t2)−n(t1)\Delta n_{i}^{\mathrm{obs}}=n(t_{2})-n(t_{1}), computed from the two SGP4 mean elements [43]; it is a change in orbital angular rate, not in density. Following the energy-dissipation approach of [32], this change is predicted by integrating the drag power along the SGP4 trajectory, written as a sum over sample points kk between t1t_{1} and t2t_{2}:

Δnipred=Bi∑kwkρ(xk,tk),wk=3an2μ∣vrel∣(vrel⋅v)Δtk,(3)\Delta n_{i}^{\mathrm{pred}}=B_{i}\sum_{k}w_{k}\rho(\mathbf{x}_{k},t_{k}),\qquad w_{k}=\frac{3an}{2\mu}\lvert\mathbf{v}_{\mathrm{rel}}\rvert(\mathbf{v}_{\mathrm{rel}}\cdot\mathbf{v})\Delta t_{k}, \tag*{(3)}

where v\mathbf{v} is the orbital velocity, vrel\mathbf{v}_{\mathrm{rel}} the velocity relative to a co-rotating atmosphere, Δtk\Delta t_{k} the sample spacing, and Bi=CDA/mB_{i}=C_{D}A/m the ballistic coefficient. The weights wkw_{k} depend only on the trajectory and are computed once in advance. The density at each sample point is given by the model of Equation (1), so that, substituting Equation (1) into Equation (3),

Δnipred=Bi∑kwkρMSIS(xk,tk)exp⁡[cd(k)+fθ(xk,zd(k))],(4)\Delta n_{i}^{\mathrm{pred}}=B_{i}\sum_{k}w_{k}\rho_{\mathrm{MSIS}}(\mathbf{x}_{k},t_{k})\exp\left[c_{d(k)}+f_{\theta}(\mathbf{x}_{k},\mathbf{z}_{d(k)})\right], \tag*{(4)}

where d(k)d(k) is the day of sample kk. Equation (4) is an explicit, differentiable function of every unknown: the network weights θ\theta, the daily states (cd,zd)(c_{d},\mathbf{z}_{d}) and the ballistic coefficients BiB_{i}. All of them are therefore estimated together by minimizing a single loss,

L=wTLE∑iκiσi2∑jh(Δnijobs−Δnijpred)+∑mh(ln⁡ρmobsρMSIS,m−δm)+R,(5)\mathcal{L}=w_{\mathrm{TLE}}\sum_{i}\frac{\kappa_{i}}{\sigma_{i}^{2}}\sum_{j}h\left(\Delta n_{ij}^{\mathrm{obs}}-\Delta n_{ij}^{\mathrm{pred}}\right)+\sum_{m}h\left(\ln\frac{\rho_{m}^{\mathrm{obs}}}{\rho_{\mathrm{MSIS},m}}-\delta_{m}\right)+\mathcal{R}, \tag*{(5)}

where jj runs over the TLE intervals of object ii, mm over accelerometer density samples, hh is the robust Huber function, σi\sigma_{i} is the noise level of object ii (estimated from its residual scatter), κi\kappa_{i} is a weight for the calibration objects (§ 3.3), and wTLE=0.3w_{\mathrm{TLE}}=0.3 balances the two data types. The gradients of L\mathcal{L} are obtained by automatic differentiation (PyTorch; [31]) and the minimization uses the Adam optimizer ([23]). Differentiable propagation has previously been demonstrated on simulated orbits ([21, 1]); here it is applied to real tracking data at scale.

Because SGP4 mean elements remove the main zonal-harmonic effects, the remaining non-drag contributions to Δnobs\Delta n^{\mathrm{obs}} are small for the selected objects; they are treated as noise by the per-object error model and screening described below. We checked Equation (3) against full numerical integration (DOP853, J2–J6 gravity plus drag) for 114 intervals stratified over 32 classes of perigee altitude (285–2000 km), eccentricity and object type, including intervals with Kp≥6K_{p}\geq6. The ratio of predicted to integrated Δn\Delta n is 1.027 (10th–90th percentile: 1.015–1.039), equal within 0.4% for storm and quiet intervals and across perigee, eccentricity and object type; the constant offset is absorbed by the fitted BiB_{i} (and biases the fixed-BB spheres, for which the ratio is 1.012–1.013, by about 1.2%). On real data, the TLEs of the accelerometer satellites themselves (1,443 intervals) agree with TU Delft densities integrated along the same samples to 5–7% per interval, with no dependence on geomagnetic activity or interval length.

Observations

TLE drag. Each year we select up to 400 rocket bodies and 400 debris objects from the Space-Track catalog, stratified by perigee between 250 and 800 km. For each object, consecutive TLE pairs (at least one day apart) are merged into longer intervals until the drag signal is clearly larger than the TLE noise. Specifically, we require that the change in mean motion expected from NRLMSISE-00 over the interval be at least five times the noise of a difference of two TLE mean motions, the latter estimated from the scatter of successive TLEs of that object. If more than 2% of an object’s intervals still show no decay (Δn≤0\Delta n\leq0), the target length is doubled, up to 60 days. Intervals are therefore longer when drag is weak, at high altitude and low solar activity. Each object has one free BiB_i per year with no prior; its drag coefficient is computed per sample from a physical gas–surface model (diffuse reflection with SESAM accommodation; [37, 34]), so that changes of CDC_D with altitude and composition are modeled and BiB_i effectively estimates a time-averaged area-to-mass ratio. For tumbling debris this average is taken over intervals much longer than the tumbling periods. Residual changes of BiB_i within a year appear as excess scatter, which raises σi\sigma_i and lowers the weight of that object in Equation (5).

Objects and intervals are screened in three ways: objects are excluded if their median BB is not positive, if more than 20% of their intervals imply a negative BB, or if more than 5% still show no decay at the maximum interval length; intervals are rejected if their implied BB differs from the object median by more than a factor of four; and residuals beyond five times σi\sigma_i are clipped. Together with the Huber function, this screening also removes manoeuvres and mis-tagged objects. After screening, 586–802 passive objects per year are used (Figure 1 and Figure S4).

NDM density at 490 km before and during the May 2024 storm, with GRACE-FO and Swarm A orbital tracks

Figure 1. NDM density at 490 km before ((a), 9 May 2024 03:00 UT, KpK_p 1.3) and during ((b), 11 May 2024 03:00 UT, KpK_p 8.3) the May 2024 storm. Tracks: one orbit of GRACE-FO and Swarm A, colored by the measured density on the same scale. Dots: the 668 TLE objects constraining NDM on that day; filled dots are within ±50\pm50 km of the slice, open dots at other altitudes (223–992 km).

Figure S4. Data used by NDM (blue), available but not used (grey), used for validation only (violet), and comparison models (orange, yellow), 2000–2026; TLE drag intervals used per month; monthly mean F10.7.

Calibration objects. Five spheres with known area-to-mass ratio (LCS-4 and Kosmos 1263, 1508, 2292 and 482; [28]) have BB fixed, so their decay directly measures the absolute density scale. Because these five objects contribute only a small fraction of the intervals, which number more than 10510^5 per year for the free objects, each of their intervals is given 30 times the normal weight (κi=30\kappa_i = 30 in Equation (5); κi=1\kappa_i = 1 otherwise); without this the calibration would be lost among the free objects. These spheres fix the absolute scale where no accelerometer is available. Other geodetic spheres (Larets, Starlette, Stella) are left free and serve as a consistency check.

Accelerometer densities. TU Delft densities from CHAMP, GRACE A/B, GOCE, Swarm A/B/C and GRACE-FO ([39, 44, 38]) enter a Huber loss on ln⁡(ρobs/ρMSIS)−δ\ln(\rho_{\mathrm{obs}}/\rho_{\mathrm{MSIS}})-\delta at their published scale. This scale lies 4–9% below the sphere scale, a difference consistent with differing drag-coefficient conventions; where accelerometers are available the product therefore follows the TU Delft scale, and where they are not it relaxes toward the sphere scale (§ 5.2). Every year from 2003 to 2025 has at least three accelerometer satellites; 2026 is reconstructed from TLEs alone and is flagged as such in the product.

Figure 1 illustrates the result: on 11 May 2024 the density at 490 km increases several-fold within a day, and the reconstruction agrees with the GRACE-FO and Swarm A tracks while the ∼670\sim670 TLE objects that constrain it are spread over all longitudes, latitudes and altitudes.

Validation Design

Because NDM assimilates every accelerometer satellite, its own fit to those satellites says little about accuracy. We therefore use two independent tests. (1) Withheld satellites: for each year, NDM is retrained six times with identical configuration, each time leaving out one accelerometer group (CHAMP; GRACE A/B; GOCE; Swarm A/C; Swarm B; GRACE-FO), and is scored against the withheld densities (60 satellite-years, 2003–2025). These retrainings are used only for evaluation and never enter the product. The same experiment is repeated with the TLEs removed (accelerometers only) and with the accelerometers removed (TLEs and spheres only) to isolate the contribution of each data source. (2) Spire GNSS orbits: precise orbits of 16–34 Spire CubeSats per year (November 2019–July 2025, 87,727 six-hour arcs, about 280–650 km) are never used in training. For each density model held fixed, a numerical propagator (with GOCO06s gravity; [24]) fits each satellite’s ballistic coefficient on odd days; on even days the arcs are re-fitted with that coefficient fixed, and the remaining 6-h 3D position error measures how well the model reproduces the drag. Because BB is fitted, this test is insensitive to a constant density scale factor: it probes the spatial and temporal variability that matters for orbit prediction, not the absolute scale. It is important to emphasize that this is not a forecast test. The density used to propagate each arc is the reanalysis itself, built from data up to and beyond that time, and the test asks how consistent the reconstructed field is with orbits that played no part in building it.

For density accuracy we use a single metric, ϵ\epsilon, the root mean square of ln⁡(ρobs/ρmodel)\ln(\rho_{\mathrm{obs}}/\rho_{\mathrm{model}}) over 94-min orbit averages. Because it is expressed in natural-log units, it is close to the rms fractional error when it is small, so that ϵ=0.10\epsilon= 0.10 means the model is typically within about 10% of the observations. We also split ϵ\epsilon into a bias (the mean of the log ratio, a systematic scale offset) and a scatter (its standard deviation, the orbit-to-orbit variability).

Each withheld-satellite experiment, one accelerometer group left out for one calendar year, gives one test case, which we call a satellite-year. There are 60 such cases, and we report the median of ϵ\epsilon over them. Their 95% intervals are obtained by bootstrap, resampling whole weeks so that the correlation between consecutive days is preserved.

For Spire, the models are compared in pairs: for each satellite, we take the ratio of its position error with one model to its error with another, and report the median ratio over satellites, with 95% intervals from resampling the satellites. When results are grouped by geomagnetic activity, a further complication arises: all satellites fly through the same storms, so satellites are not independent within a KpK_p bin. For these comparisons we therefore resample whole geomagnetic events, defined as runs of consecutive days in the same daily-maximum-KpK_p bin (5,000 resamples), and give no interval for bins with fewer than five events.

NDM is compared with four baselines: NRLMSISE-00, NRLMSIS 2.1, the SET HASDM database [42] extended to 2025, and, from July 2023, the operational WAM-IPE model of NOAA SWPC [16].

Results

Density on Satellites Never Seen in Training

Over the 60 withheld satellite-years the median error is ϵ=0.078\epsilon= 0.078 for NDM, 0.101 for HASDM and 0.310 for NRLMSISE-00 (Figure 2a, Table 2): 23% below HASDM and a factor of four below the empirical model. NDM outperforms HASDM in 45 of 60 cases, and in every year from 2009 onward except 2011 and 2020, where the two are tied. By daily maximum KpK_p (Figure 2b) the median error of NDM grows only from 0.074 to 0.095, while that of HASDM grows from 0.096 to 0.161.

Model2003–2012 (22)2013–2025 (38)All (60)
NDM (TLEs + spheres + accelerometers)0.155 / 0.2230.066 / 0.0990.078 / 0.180
NDM, accelerometers only0.092 / 0.3310.083 / 0.1460.084 / 0.228
NDM, TLEs + spheres only0.178 / 0.2600.144 / 0.2120.151 / 0.236
HASDM0.176 / 0.2110.088 / 0.1210.101 / 0.206
NRLMSISE-000.320 / 0.5300.310 / 0.6420.310 / 0.602

Table 2. Density Error on Withheld Satellites by Data Source1

Figure 2. Density error on accelerometer satellites withheld from training, ϵ=rms⁡\epsilon= \operatorname{rms} of ln⁡(ρobs/ρmodel)\ln(\rho_{\mathrm{obs}}/\rho_{\mathrm{model}}) on 94-min orbit averages. (a) Median over the satellites withheld in each year, 2003–2025. (b) Median over the 60 satellite-years, by daily maximum KpK_p. Blue: NDM retrained without the satellite; orange: HASDM; green: NRLMSISE-00.

The decomposition shows where this advantage comes from. The median bias is −0.034-0.034 for NDM and −0.074-0.074 for HASDM, but the median scatter is 0.063 for NDM and 0.052 for HASDM, which has the smaller scatter in 49 of 60 cases; on storm days (Kp≥6K_p \ge6) since 2013 the scatter is equal (0.074) while the biases are +0.009 and −0.085-0.085. NDM’s lower total error therefore comes mainly from its scale, learned from the other TU Delft satellites. The HASDM offset of about 7% matches the 4–9% offset between the TU Delft densities and the calibration spheres, so it may largely reflect a drag-coefficient convention rather than a model error; HASDM reproduces orbit-to-orbit variability slightly better, consistent with the Spire results below.

The reanalysis also quantifies the bias of NRLMSISE-00 over two solar cycles. The global-mean correction at 400–500 km reaches −50-50 to −57-57% during the 2008–2009 and 2019–2020 minima (Figure S2), consistent with the record-low densities reported for the 2008 minimum [14].

Figure S2. Global-mean correction of NRLMSISE-00 by the NDM reanalysis at 300–800 km (27-day running mean, top) and 81-day mean F10.7 (bottom), October 2003–October 2026. The reanalysis lowers NRLMSISE-00 by 50–57% at 400–500 km during the 2008–2009 and 2019–2020 solar minima.

What the Tracked Objects Contribute

Table 2 separates the data sources. From 2013 to 2025, adding the TLEs to the accelerometers lowers the median error by 20% (0.083 to 0.066) and the 90th percentile from 0.146 to 0.099. More importantly, the TLEs remove the failures of accelerometer-only extrapolation, which reach ϵ=0.83\epsilon= 0.83 (Swarm A/C withheld in 2018) and 0.57 (2020); with TLEs the largest error after 2012 is 0.14. Trained on TLEs and spheres alone, with no accelerometer data, NDM reaches a median of 0.15, half the error of NRLMSISE-00; this is the configuration of the 2026 product, for which we expect a median error of 0.11–0.14 (the TLE-only values in 2022–2025). On Spire orbits the TLE-only version reduces the NRLMSISE-00 error by 20–44% in 2022–2025 but by less than 10% at the 2019–2021 solar minimum.

From 2003 to 2012 the picture is reversed: the TLEs raise the median error from 0.092 to 0.155, although they also remove the two worst accelerometer-only failures (0.55 and 0.35). In those years the accelerometers fly at only two or three altitudes, often 100–200 km apart, near solar minimum, where the TLE drag signal is weakest; the daily latent state then absorbs TLE noise, which increases day-to-day scatter at altitudes without accelerometer data. The error is smallest when a training satellite lies within 100 km in altitude of the evaluated orbit (Figure S1).

Figure S1. Density error on withheld accelerometer satellites, by altitude of the withheld satellite, annual mean F10.7, and altitude gap to the nearest accelerometer satellite remaining in training. NDM is outperformed by HASDM only when the gap exceeds 100 km (mostly GRACE withheld in 2004–2012). HASDM data courtesy of Space Environment Technologies.

Two further tests, both in the production configuration, address whether the observations, rather than the network, determine the spatial structure (Text S1). In the first, a synthetic experiment, we prescribe a known density field made of separate components (a daily global level, an altitude gradient, a latitude pattern, a diurnal variation, small-scale structure, and 3-day and 6-h oscillations), generate the TLE observations it would produce with the real 2024 sampling, noise and drag operator, and train NDM on them. For each component we report the gain, the ratio of recovered to true amplitude (1 means fully recovered, 0 not recovered at all), and, where useful, the correlation between recovered and true patterns. The TLEs alone recover the daily global level (gain 1.00, correlation >0.99), the altitude gradient (gain 0.98, correlation 0.97), the latitude structure (pattern correlation 0.81–0.87, but with an amplitude 18–29% too large), the diurnal variation (gain 1.03–1.08, phase within 0.2 h below 700 km) and 3-day oscillations (gain 0.89–1.01). They do not recover small-scale latitude–local-time patterns or 6-h oscillations (gains near zero). Overall, the recovered spatial pattern correlates with the truth at 0.86–0.89. A simpler correction made of a daily scale factor with a daily altitude slope recovers the altitude gradient but no horizontal structure (grid error 0.10, against 0.065 for NDM), and without the daily state the 3-day variability is recovered with a gain of only 0.28–0.35.

In the second test, on real data with GRACE-FO withheld (Table S1), replacing the neural correction by a daily global scale factor raises ϵ\epsilon from 0.050 to 0.100 in 2023 and from 0.046 to 0.074 in 2024, and on storm days it more than doubles it. Removing the daily state raises ϵ\epsilon by only about 10% (to 0.055 and 0.050), and on 2024 storm days not at all. Spatially resolved estimation thus provides most of the skill. The daily state is needed mainly for multi-day variability that the geomagnetic indices do not explain; structure at scales below the resolution above is inherited from NRLMSISE-00 and the network.

Correction202320242023 storm days2024 storm days
Full: cd+fθ(x,zd)c_d + f_{\theta}(\mathbf{x}, z_d)0.0500.0460.0660.061
Daily scale factor: cdc_d0.1000.0740.1570.143
Daily scale and altitude slope0.1000.0740.1580.143
No daily state: fθ(x)f_{\theta}(\mathbf{x})0.0550.0500.0700.056
HASDM0.0750.0820.1060.111
NRLMSISE-000.2440.3400.1670.388

Table S1. Table S1. Error ϵ\epsilon on Withheld GRACE-FO for Reduced Forms of the Correction (Production Configuration)1

Five Major Storms of 2023–2024

Figure 3 shows five storms observed by GRACE-FO, with GRACE-FO withheld from NDM (Swarm and all TLEs retained). All errors in this section are values of ϵ\epsilon computed over a 7-day window around each storm, and biases are in the same natural-log units. NDM reaches ϵ=0.046\epsilon= 0.046–0.062, against 0.075–0.110 for HASDM, 0.14–0.56 for NRLMSISE-00 and 0.20–0.86 for WAM-IPE. For the May 2024 (Gannon) storm, ϵ\epsilon is 0.51 for NRLMSISE-00 and 0.56 for WAM-IPE, while NDM follows the observed rise and recovery with ϵ=0.062\epsilon= 0.062. NDM is nearly unbiased (bias −0.02 to −0.04), whereas HASDM has a bias of −0.07 to −0.09, meaning that it overestimates the observed density by 7–9%, again comparable to the offset between the TU Delft and sphere scales. The NDM main phase is often slightly late or smoothed, a consequence of the daily state and the 1–3-day TLE intervals (§ 6).

Figure 3. Five major storms observed by GRACE-FO (94-min orbit averages, ±1.5\pm1.5 days around each peak), with GRACE-FO withheld from NDM. Black: observed (TU Delft). Seven-day errors ϵ\epsilon for NDM / HASDM / NRLMSISE-00 / WAM-IPE: (a) 5 Nov 2023, 0.054 / 0.086 / 0.14 / 0.20; (b) 24 Mar 2024, 0.061 / 0.080 / 0.24 / 0.46; (c) 11 May 2024, 0.062 / 0.102 / 0.51 / 0.56; (d) 12 Aug 2024, 0.048 / 0.075 / 0.34 / 0.64; (e) 10 Oct 2024, 0.046 / 0.110 / 0.56 / 0.86.

Independent Test on Spire GNSS Orbits

On Spire orbits NDM reduces the 6-h position error relative to NRLMSISE-00 in every year from 2019 to 2025 (Table 1). The improvement is statistically significant in every year: the 95% interval of the median ratio of the two errors, taken satellite by satellite, lies entirely below one. Relative to HASDM, the result depends on solar activity and altitude: NDM is better by 4–5% during the 2019–2021 solar minimum, tied in 2023, and behind by 5–16% in 2022, 2024 and 2025. The HASDM advantage lies mainly at 550–650 km and 400–450 km, outside the altitude band of the training accelerometers. In the WAM-IPE period (Figure 4a) NDM reduces the error by 38–69%; in 2024 WAM-IPE is no better than NRLMSISE-00. The ballistic coefficients fitted with NDM or HASDM scatter by 6–8% from arc to arc in 2022–2025, against 15–17% with NRLMSISE-00, a direct measure of improved orbit predictability. Since BB is fitted, these results concern the variability of the field, not its absolute scale.

YearArcsNRLMSISE-00NRLMSIS 2.1HASDMNDMNDM vs HASDM
2019 (Nov–Dec)3,2420.410.400.410.37−5% (n.s.)
202012,5300.610.580.520.48−4.5% [−6, −3]
2021 (to Nov)15,7700.660.640.540.51−4% [−7, −2]
202226,2892.622.581.241.29+5.5% [+2, +8]
202321,76511.6911.174.895.06+0.1% [−2, +6]
20243,39316.4520.019.6810.10+16% [+7, +27]
2025 (to Jul)4,7382.362.401.541.62+5% [+2, +15]

Table 1. Spire 6-h 3D Position Error (m) With Fixed Ballistic Coefficient1

Independent test on Spire GNSS orbits

Figure 4. Independent test on Spire GNSS orbits (6-h arcs, ballistic coefficient fixed from other days). (a) Median 6-h position error in the WAM-IPE period (23 Jul 2023–30 Jul 2025), same arcs for all models. (b–d) Paired median change in position error of NDM relative to HASDM, by daily maximum KpK_p, for (b) 2022–2025, (c) 2019–2021 and (d) 2019–2025; negative values mean NDM is better. Error bars are 95% intervals from resampling whole geomagnetic events; open symbols mark bins with fewer than five events, for which no interval is given. © Spire Global, Inc. 2019–2025.

By geomagnetic activity (Figure 4b–Figure 4d), NDM and HASDM agree to within a few percent for Kp<7K_p < 7. In 2022–2025, HASDM, which assimilates radar tracking, has errors 4–7% lower, a difference that is significant for Kp<5K_p < 5. On days with Kp≥7K_p \ge7, the median error of NDM is 12% lower than that of HASDM over 2019–2025, and its error relative to NRLMSISE-00 is −59% against −53% for HASDM (Figure S3). This category, however, contains only 13 storm events with test arcs, two of them in 2019–2021 (Table S2). Resampling whole events gives a 95% interval for the difference that ranges from NDM being 35% better to 2% worse: NDM is better in 8 of the 13 events, but HASDM is better in the two largest (May and October 2024). Because this interval includes zero, the data are consistent with no real difference between the two models in severe storms; the 12% advantage of NDM is the best single estimate, but with only 13 storms it cannot be considered established.

StartDaysMax KpK_pStartDaysMax KpK_p
2021-05-1217.02024-05-1039.0
2021-11-0417.72024-06-2817.7
2022-04-1016.72024-08-0417.0
2022-08-1716.72024-08-1218.0
2023-02-2716.72024-09-1217.0
2023-03-2328.02024-09-1717.3
2023-04-2328.32024-10-0838.7
2023-11-0517.32025-01-0118.0
2023-12-0116.72025-04-1617.7
2024-04-1917.02025-06-0137.7
2024-05-0216.72025-06-1316.7

Table S2. Table S2. Storms With Daily Maximum Kp≥7−K_p \ge7- During the Spire Coverage (November 2019–July 2025)2

Figure S3. Spire 6-h position error relative to NRLMSISE-00 (median change, 95% bootstrap interval; negative = better), by altitude, local time of the ascending node, arc maximum KpK_p and daily F10.7, for NDM, HASDM, NRLMSIS 2.1 and JB2008. JB2008 was run as configured in the pyatmos package; its setup has not been independently verified. © Spire Global, Inc. 2019–2025; data obtained through NASA CSDA.

Discussion and Conclusions

NDM shows that the orbital decay of a few hundred passive objects per year, combined with a few known-shape spheres and the accelerometer record, constrains the global 3D thermospheric density with skill comparable to the operational product that assimilates non-public radar data: lower total error on withheld accelerometers, mainly through a smaller bias, and similar performance on independent orbits.

The key methodological step is to fit the density model directly to the observed orbital decay. In the conventional use of TLEs ([32, 15, 10]), the decay of each object over each interval is first converted into a single density value: an average weighted along the orbit, which depends on an assumed ballistic coefficient and is assigned to one location, usually near perigee, through an assumed altitude profile. Errors in the assumed coefficient and profile, and the loss of information about where along the orbit the drag occurred, then enter every subsequent step. In NDM no such intermediate densities exist. The fully differentiable chain from the network parameters through Equation (4) to the observed changes in mean motion compares each observation with exactly the quantity it measures, the drag integrated along the actual trajectory. This allows the density field, the daily state and ∼700 ballistic coefficients to be estimated jointly and consistently from heterogeneous data.

The second choice that makes NDM possible is that the neural correction is spatially resolved. In contrast to the few global drivers estimated by most assimilation schemes ([5, 41, 28] (Foorotan et al. 2023)), NDM learns latitude, local-time and altitude structure from objects in many orbital planes, which a daily global scale factor cannot reproduce (§ 5.2).

What observation-constrained should be distinguished from what is model-driven. The daily level, the altitude gradient, the latitude pattern (although not its exact amplitude) and the diurnal variation below about 700 km, at time scales of a day or longer, are constrained by the data. Sub-daily storm evolution and small-scale structure come from the geomagnetic drivers, the network’s learned response to them and NRLMSISE-00. Storm main phases are therefore under-estimated and lagged by hours (cf. [8]); shorter TLE intervals did not remove this lag, and resolving it will require precise orbits, for example through a full differentiable orbit propagator applied to GNSS-tracked satellites.

The current version has further limitations. (1) In 2003–2012 the TLEs increase the median error relative to accelerometers alone, and away from accelerometer altitudes near solar minimum the TLE noise increases day-to-day scatter. (2) The absolute scale is uncertain by about ±10%, from the area-to-mass ratios of the calibration spheres and the modeled drag coefficient, and the TU Delft and sphere scales differ by 4–9%. (3) The yearly networks produce seams at year boundaries with a near-zero mean jump and an rms jump of 0.02–0.05, about twice the seed spread; the archived product will include the ±15-day overlaps so that users can blend across boundaries. (4) The 2026 field is constrained by TLEs alone. Without accelerometers the scale relaxes toward the sphere scale: in 2022–2025 the TLE-only fields lie 0.02–0.12 (ln units, increasing with altitude from 300 to 700 km) above the product, so 2026 is expected to lie above the TU Delft level by about this amount, with a median error of 0.11–0.14 based on the retrospective TLE-only validation. (5) Uncertainty quantification is not yet included. The seed spread (0.006– 0.016) is much smaller than the validation errors and is not a usable uncertainty. Until a full error model is available, the validation errors by altitude gap, solar activity and KpK_p (Figure 2 and Figure S1) provide an empirical guide: typically 0.04–0.10 within 100 km of an accelerometer satellite after 2012, and up to 0.15–0.30 for upward extrapolation at solar minimum. (6) In the synthetic test the amplitude of the latitude structure is overestimated by 18–29% (13–18% without noise), so the overshoot comes partly from the sampling and not only from the noise.

In conclusion, the thermosphere has so far lacked what weather forecasting gained from reanalyses: a long, continuous and observation-constrained record of the 3D state that anyone can reproduce. NDM provides such a record for 2003–2026. It combines the decay of tracked debris and rocket bodies with a few calibration spheres and the accelerometer missions, and it reaches an accuracy on unseen satellites comparable to that of HASDM and four times better than that of NRLMSISE-00. Because it relies only on public observations, it can be extended in near-real time as new TLEs and new missions become available. We expect it to be useful in three ways: as a reference for orbit determination and collision avoidance, as a benchmark for physics-based and empirical models, and as a training set for the next generation of machine-learning forecasts of thermospheric density (e.g., [25, 45]).

Open Research

TLEs are available from Space-Track.org (https://www.space-track.org, free registration). Accelerometer and GNSS-derived densities of CHAMP, GRACE, GOCE, Swarm and GRACE-FO are available from TU Delft (http://thermosphere.tudelft.nl). Geomagnetic and solar indices are from GFZ Potsdam ([26]). Spire GNSS orbits were obtained through the NASA Commercial Satellite Data Acquisition (CSDA) Program and are available to eligible users from NASA CSDA. SET HASDM densities are available for scientific use on request from Space Environment Technologies ([42]). WAM-IPE output is distributed by NOAA SWPC. NRLMSISE-00 and NRLMSIS 2.1 were evaluated with the open-source pymsis package. The NDM reanalysis (daily 3D grids, 2003–2026) and the code used to produce it will be deposited at [Zenodo DOI to be added upon acceptance] under an open licence, together with the configuration files and scripts needed to reproduce the withheld-satellite and Spire experiments. The archived fields include the ±15-day overlaps between yearly networks and a flag marking the TLE-only period (2026).

Acknowledgments

This work utilized data made available through the NASA Commercial Satellite Data Acquisition (CSDA) Program; Spire data © Spire Global, Inc. 2019–2025. The SET HASDM density data are provided for scientific use courtesy of Space Environment Technologies. We thank TU Delft for the accelerometer and POD densities, NOAA SWPC for WAM-IPE output, GFZ Potsdam for the geomagnetic and solar indices, the 18th Space Defense Squadron and Space-Track.org for TLEs, and Natural Earth for coastlines.

Supporting Information for “The Neural Density Model: An Open 3D Thermospheric Density Reanalysis for 2003–2026 From Tracked Space Debris”

Enrico Camporeale

supporting_information.pdf

Introduction Text S1 and Table S1 describe two tests, in the production configuration, of which parts of the density field are constrained by the TLE observations. Table S2 lists the severe storms in the Spire validation period. Figures S1–S4 give the breakdown of the density and Spire validation by altitude, solar activity, local time and geomagnetic activity, the global-mean correction of NRLMSISE-00 by the NDM reanalysis over 2003–2026, and the data used in each period. All density errors are the rms of ln⁡(ρobs/ρmodel)\ln(\rho_{\mathrm{obs}/\rho_{\mathrm{model}}}) on 94-min orbit averages, computed with NDM retrained without the evaluated satellite.

Text S1. Identifiability tests. Both tests use the production configuration. This means the production flags and regularization of the year, all TLE objects, and the five calibration objects at their fixed area-to-mass ratios. The drag coefficient is computed per sample with the same operator as in training.

Synthetic recovery. Synthetic mean-motion changes were generated on the real screened TLE intervals of 2024 from a known correction δtrue\delta_{\mathrm{true}}. The ballistic coefficients were taken from the production run, and the per-object noise from that run’s TLE residuals. The truth contained a daily global level (random walk, step 0.04), an altitude gradient (0.15 per 200 km, modulated by ±50% over 27 days), a latitude P2P_2 term (0.10), a diurnal term (cos⁡(LST−14 h)cos⁡ϕ,0.10)(\cos(\mathrm{LST}-14\ \mathrm{h})\cos\phi, 0.10), a small-scale term (sin⁡3 LSTsin⁡2ϕ,0.05)(\sin3\ \mathrm{LST}\sin2\phi, 0.05), a 3-day oscillation (0.06) and a 6-h oscillation (0.05). NDM was trained on TLEs alone, with realistic noise and, for comparison, without noise. Over 300–800 km it recovered:

  • the daily level with gain 1.00 (correlation ≥0.997);

  • the altitude gradient with gain 0.98 (slope correlation 0.97);

  • the latitude structure with pattern correlation 0.81–0.87, but an amplitude 18–29% too large (13–18% without noise);

  • the diurnal term with gain 1.03–1.08, pattern correlation 0.75–0.77, and a phase error ≤0.2 h below 700 km (0.6 h at 800 km);

  • the 3-day oscillation with gain 0.89–1.01.

The small-scale and 6-h terms were not recovered (gain ≈0). With the global mean removed at each time step, the recovered spatial pattern correlates with the truth at 0.86–0.89 and explains 63–68% of its variance; the unrecoverable small-scale term limits this value. Three reduced corrections were trained on the same synthetic data:

  • A daily scale factor with a daily altitude slope, δ=cd+sd(h−500 km)/200 km\delta= c_d+s_d\left(h-500\ \mathrm{km}\right)/200\ \mathrm{km}, recovers the altitude gradient (gain 1.12) but no horizontal structure; its grid error is 0.10, against 0.065 for NDM.

  • A daily scale factor alone recovers neither.

  • Without the daily state, the spatial pattern is recovered but the 3-day oscillation is not (gain 0.28–0.35).

Reduced corrections on real data. Table S1 compares the full correction with the three reduced forms, using the production withheld-satellite protocol: GRACE-FO withheld for the whole year, with Swarm A/B/C and all TLEs kept. Because the altitude slope is defined about 500 km and GRACE-FO flies near 490 km, the scale-plus-slope variant cannot differ from the scale-only variant on this test; it is evaluated by the synthetic experiment above.

References

  1. [1]Acciarini, G., Baydin, A. G., & Izzo, D. (2024). Closing the gap between SGP4 and high-precision propagation via differentiable programming. (arXiv:2402.04830) doi: 10.48550/arXiv.2402.04830
  2. [2]Berger, T. E., Dominique, M., Lucas, G., Pilinski, M., Ray, V., Sewell, R., et al. (2023). The thermosphere is a drag: The 2022 Starlink incident and the threat of geomagnetic storms to low Earth orbit space operations. Space Weather, 21(3), e2022SW003330. doi: 10.1029/2022SW003330DOI
  3. [3]Berger, T. E., Holzinger, M. J., Sutton, E. K., & Thayer, J. P. (2020). Flying through uncertainty. Space Weather, 18(1), e2019SW002373. doi: 10.1029/2019SW002373DOI
  4. [4]Bowman, B. R., Tobiska, W. K., Marcos, F. A., Huang, C. Y., Lin, C. S., & Burke, W. J. (2008). A new empirical thermospheric density model JB2008 using new solar and geomagnetic indices. In AIAA/aas astrodynamics specialist conference and exhibit (p. AIAA 2008-6438). Honolulu, HI. doi: 10.2514/6.2008-6438DOI
  5. [5]Brandt, D. A., Bussy-Virat, C. D., & Ridley, A. J. (2020). A simple method for correcting empirical model densities during geomagnetic storms using satellite orbit data. Space Weather, 18, e2020SW002565. doi: 10.1029/2020SW002565
  6. [6]Bruinsma, S., Sutton, E., Solomon, S. C., Fuller-Rowell, T., & Fedrizzi, M. (2018). Space weather modeling capabilities assessment: Neutral density for orbit determination at low Earth orbit. Space Weather, 16(11), 1806–1816. doi: 10.1029/2018SW002027DOI
  7. [7]Cheng, C.-C., Fuller-Rowell, T., Sutton, E. K., Fang, T.-W., Liu, J.-Y., & Weimer, D. R. (2024). Thermospheric neutral density data assimilation system based on the Whole Atmosphere Model during the November 2003 storm. Space Weather, 22, e2024SW004010. doi: 10.1029/2024SW004010DOI
  8. [8]Constant, C., Brownhall, I., Aruliah, A., Ziebart, M., & Bhattarai, S. (2026). Empirical assessment of storm-time thermospheric density inversion methods from LEO POD data. Earth and Space Science, 13, e2025EA004914. doi: 10.1029/2025EA004914DOI
  9. [9]Dietrich, N., & Matsuo, T. (2026). Quantifying storm-time neutral density uncertainties using a physics-based particle filter framework. The Journal of the Astronautical Sciences, 73, 49. doi: 10.1007/s40295-026-00595-xDOI
  10. [10]Doornbos, E., Klinkrad, H., & Visser, P. (2008). Use of two-line element data for thermosphere neutral density model calibration. Advances in Space Research, 41(7), 1115–1122. doi: 10.1016/j.asr.2006.12.025
  11. [11]Doornbos, E., van den IJssel, J., Lühr, H., Förster, M., & Koppenwallner, G. (2010). Neutral density and crosswind determination from arbitrarily oriented multiaxis accelerometers on satellites. Journal of Spacecraft and Rockets, 47(4), 580–589. doi: 10.2514/1.48114DOI
  12. [12]Emmert, J. T. (2015). Thermospheric mass density: A review. Advances in Space Research, 56(5), 773–824. doi: 10.1016/j.asr.2015.05.038DOI
  13. [13]Emmert, J. T., Jones Jr., M., Siskind, D. E., Drob, D. P., Picone, J. M., Stevens, M. H., et al. (2022). NRLMSIS 2.1: An empirical model of nitric oxide incorporated into MSIS. Journal of Geophysical Research: Space Physics, 127(10), e2022JA030896. doi: 10.1029/2022JA030896DOI
  14. [14]Emmert, J. T., Lean, J. L., & Picone, J. M. (2010). Record-low thermospheric density during the 2008 solar minimum. Geophysical Research Letters, 37, L12102. doi: 10.1029/2010GL043671DOI
  15. [15]Emmert, J. T., Picone, J. M., & Meier, R. R. (2008). Thermospheric global average density trends, 1967–2007, derived from orbits of 5000 near-Earth objects. Geophysical Research Letters, 35, L05101. doi: 10.1029/2007GL032809DOI
  16. [16]Fang, T.-W., Kubaryk, A., Goldstein, D., Li, Z., Fuller-Rowell, T., Millward, G., et al. (2022). Space weather environment during the SpaceX Starlink satellite loss in February 2022. Space Weather, 20(11), e2022SW003193. doi: 10.1029/2022SW003193DOI
  17. [17]Forootan, E., Farzaneh, S., Kosary, M., Borries, C., Kodikara, T., & Schumacher, M. (2023). Predicting global thermospheric neutral density during periods with high geomagnetic activity. Scientific Reports, 13, 20322. doi: 10.1038/s41598-023-47440-xDOI
  18. [18]Forootan, E., Farzaneh, S., Kosary, M., Schmidt, M., & Schumacher, M. (2021). A simultaneous calibration and data assimilation (C/DA) to improve NRLMSISE00 using thermospheric neutral density (TND) from space-borne accelerometer measurements. Geophysical Journal International, 224(2), 1096–1115. doi: 10.1093/gji/ggaa507
  19. [19]Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049. doi: 10.1002/qj.3803DOI
  20. [20]Huang, W., Xia, S., Sun, J., Zhao, X., Lei, X., Zhao, G., & Chen, J. (2026). Calibration method of empirical atmospheric density based on physics-informed neural networks. Space Weather, 24, e2026SW005147. doi: 10.1029/2026SW005147DOI
  21. [21]Izzo, D., Acciarini, G., & Biscani, F. (2024). NeuralODEs for VLEO simulations: Introducing thermoNET for thermosphere modeling. In 29th international symposium on space flight dynamics (issfd). (arXiv:2405.19384) doi: 10.48550/arXiv.2405.19384DOI
  22. [22]Jacchia, L. G. (1970). New static models of the thermosphere and exosphere with empirical temperature profiles (SAO Special Report No. 313). Cambridge, MA: Smithsonian Astrophysical Observatory.
  23. [23]Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. In 3rd international conference on learning representations (iclr). San Diego, CA. (arXiv:1412.6980)arxiv.org/abs/1412.6980
  24. [24]Kvas, A., Brockmann, J. M., Krauss, S., Schubert, T., Gruber, T., Meyer, U., et al. (2021). GOCO06s – a satellite-only global gravity field model. Earth System Science Data, 13(1), 99–118. doi: 10.5194/essd-13-99-2021DOI
  25. [25]Licata, R. J., Mehta, P. M., Tobiska, W. K., & Huzurbazar, S. (2022). Machine-learned HASDM thermospheric mass density model with uncertainty quantification. Space Weather, 20(4), e2021SW002915. doi: 10.1029/2021SW002915DOI
  26. [26]Matzka, J., Stolle, C., Yamazaki, Y., Brönkalla, O., & Morschhauser, A. (2021). The geomagnetic Kp index and derived indices of geomagnetic activity. Space Weather, 19(5), e2020SW002641. doi: 10.1029/2020SW002641DOI
  27. [27]Mehta, P. M., & Linares, R. (2017). A methodology for reduced order modeling and calibration of the upper atmosphere. Space Weather, 15(10), 1270–1287. doi: 10.1002/2017SW001642DOI
  28. [28]Mutschler, S., Pilinski, M., Zesta, E., Oliveira, D. M., Delano, K., Garcia-Sage, K., & Tobiska, W. K. (2026). First results of a new inversion tool for thermospheric neutral mass density computations during severe geomagnetic storms. AGU Advances, 7, e2025AV002079. doi: 10.1029/2025AV002079DOI
  29. [29]Oliveira, D. M., & Zesta, E. (2019). Satellite orbital drag during magnetic storms. Space Weather, 17, 1510–1533. doi: 10.1029/2019SW002287DOI
  30. [30]Parker, W. E., & Linares, R. (2024). Satellite drag analysis during the May 2024 Gannon geomagnetic storm. Journal of Spacecraft and Rockets, 61(5), 1412–1416. doi: 10.2514/1.A36164DOI
  31. [31]Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., et al. (2019). PyTorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems (Vol. 32, pp. 8024–8035).arxiv.org/abs/1912.01703
  32. [32]Picone, J. M., Emmert, J. T., & Lean, J. L. (2005). Thermospheric densities derived from spacecraft orbits: Accurate processing of two-line element sets. Journal of Geophysical Research: Space Physics, 110(A3), A03301. doi: 10.1029/2004JA010585DOI
  33. [33]Picone, J. M., Hedin, A. E., Drob, D. P., & Aikin, A. C. (2002). NRLMSISE-00 empirical model of the atmosphere: Statistical comparisons and scientific issues. Journal of Geophysical Research: Space Physics, 107(A12), 1468. doi: 10.1029/2002JA009430DOI
  34. [34]Pilinski, M. D., Argrow, B. M., Palo, S. E., & Bowman, B. R. (2013). Semi-empirical satellite accommodation model for spherical and randomly tumbling objects. Journal of Spacecraft and Rockets, 50(3), 556–571. doi: 10.2514/1.A32348DOI
  35. [35]Ponder, B. M., Ridley, A. J., Goel, A., & Bernstein, D. S. (2023). Improving forecasting ability of GITM using data-driven model refinement. Space Weather, 21, e2022SW003290. doi: 10.1029/2022SW003290DOI
  36. [36]Ray, V., Thayer, J., Sutton, E. K., & Waldron, Z. (2024). Error assessment of thermospheric mass density retrieval with POD products using different strategies during solar minimum. Space Weather, 22, e2023SW003585. doi: 10.1029/2023SW003585DOI
  37. [37]Sentman, L. H. (1961). Free molecule flow theory and its application to the determination of aerodynamic forces (Tech. Rep. No. LMSC-448514). Sunnyvale, CA: Lockheed Missiles and Space Company.DOI
  38. [38]Siemes, C., Borries, C., Bruinsma, S., Fernandez-Gomez, I., Hladczuk, N., van den IJssel, J., et al. (2023). New thermosphere neutral mass density and crosswind datasets from CHAMP, GRACE, and GRACE-FO. Journal of Space Weather and Space Climate, 13, 16. doi: 10.1051/swsc/2023014DOI
  39. [39]Siemes, C., van den IJssel, J., Doornbos, E., Krauss, S., Pérez, J. D., Floberghagen, R., et al. (2016). Swarm accelerometer data processing from raw accelerations to thermospheric neutral densities. Earth, Planets and Space, 68, 92. doi: 10.1186/s40623-016-0474-5DOI
  40. [40]Storz, M. F., Bowman, B. R., Branson, M. J. I., Casali, S. J., & Tobiska, W. K. (2005). High accuracy satellite drag model (HASDM). Advances in Space Research, 36(12), 2497–2505. doi: 10.1016/j.asr.2004.02.020DOI
  41. [41]Sutton, E. K., Thayer, J. P., Pilinski, M. D., Mutschler, S. M., Berger, T. E., Nguyen, V., & Masters, D. (2021). Toward accurate physics-based specifications of neutral density using GNSS-enabled small satellites. Space Weather, 19, e2021SW002736. doi: 10.1029/2021SW002736DOI
  42. [42]Tobiska, W. K., Bowman, B. R., Bouwer, D., Cruz, A., Wahl, K., Pilinski, M., et al. (2021). The SET HASDM density database. Space Weather, 19(4), e2020SW002682. doi: 10.1029/2020SW002682DOI
  43. [43]Vallado, D. A., Crawford, P., Hujsak, R., & Kelso, T. S. (2006). Revisiting Spacetrack Report #3. In Aiaa/aas astrodynamics specialist conference and exhibit (p. AIAA 2006-6753). Keystone, CO. doi: 10.2514/6.2006-6753DOI
  44. [44]van den IJssel, J., Doornbos, E., Iorfida, E., March, G., Siemes, C., & Montenbruck, O. (2020). Thermosphere densities derived from Swarm GPS observations. Advances in Space Research, 65(7), 1758–1771. doi: 10.1016/j.asr.2020.01.004DOI
  45. [45]Wang, R., & Bai, X. (2026). A machine-learning-based global thermospheric density forecasting model. Space Weather, 24, e2026SW004968. doi: 10.1029/2026SW004968DOI

Paper details

vlnc:W_7adv6hy3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents