Descoberta · open engineering notebook

Orbit Weather · LEO drag & density in 3D

Fly a satellite through the thermosphere. Pick an orbit and how active the Sun is, then watch the air density around the planet, the drag on the spacecraft, and how fast it comes down. Then replay a real event: the February 2022 storm that brought down 38 new Starlink satellites, and see what the standard model expected against what happened. Everything runs in your browser, from open physics you can read below.

Orbit

Sun & storms

Spacecraft

Area is the cross-section facing the flow (a tumbling 3U averages about 0.03 m²). CD ≈ 2.2 is the usual first guess for a compact satellite.

Loading 3D…

Density here
–
Drag here
–
Altitude loss
–
Lifetime*
–
· deceleration
Artist's illustration. A generic small satellite: the orange port on the face that meets the airflow (the ram face; use "Ram face" view) is where an in-place density sensor looks. It is not any specific design. The satellite is drawn hugely larger than scale, and the shell shows density at your altitude only (bright = denser). Sun fixed to the right, equinox, orbit plane held fixed against the Sun. Density, drag and decay are computed live from our open model; not a measurement. *Lifetime holds today's Sun fixed.

Density along one orbit · the white dot is your satellite

Your settingsSolar minSolar maxStorm (F10.7 150, Ap 150)

Results

Neutral density
kg/m³ ·
Drag force
Drag acceleration
Altitude loss
Lifetime, Sun held fixed
to re-entry (≈150 km)
Ballistic coefficient
m / (CDA), kg/m²

Altitude loss and lifetime always use the mean density along your orbit.

Density vs altitude

Solar min (F10.7 70, Ap 4)Moderate (120, Ap 8)Solar max (200, Ap 15)Your settings

Replay a real event: the Starlink storm, February 2022

On 3 February 2022 a Falcon 9 put 49 new Starlink satellites into a low checkout orbit (about 210 by 335 km). A moderate geomagnetic storm arrived the next day. SpaceX said onboard GPS showed drag “up to 50 percent higher than during previous launches”. The satellites couldn't climb out, and 38 of the 49 burned up. Here is the same week, flown again with real space-weather data.

+13%How much denser a standard empirical model (the open one on this page, fed the real storm index) made the air at 210 km at midday on 4 Feb, against a quiet day. simulated
up to +50%The extra drag SpaceX saw on the satellites' own GPS, against earlier launches (SpaceX statement, 8 Feb 2022). reported
38 of 49Satellites that re-entered between 5 and 12 Feb, from Jonathan McDowell's public catalogue (32 of the dates are his estimates). observed

Altitude, day by day · lines = simulation, dots = real re-entries (hollow = estimated date)

Now

Time
Ap (storm index)
F10.7 (Sun)
Altitude

Area facing the air

The one number nobody outside SpaceX knows. Flying "edge-on, like a sheet of paper" (SpaceX's safe mode) shrinks it; a tumbling or flat satellite shows much more. Mass 297 kg and size from McDowell's catalogue; CD 2.2.

When does it come down? · same satellite, same week, four versions of the air

What would have happened

SpaceX launches low on purpose: a satellite that fails its checkout falls out of the sky within days instead of becoming debris. Healthy ones fire their thrusters and climb. So at 210 km every line above comes down within days unless the thrusters win. The question that February was whether they could win against the air the storm added. The blue line is the same week with a quiet Sun. The gap between it and the others is time the satellites lost, inside a window that was already short.

This replay has no thrust in it. It shows the air the satellites were fighting, not SpaceX's manoeuvres.

Working it backwards: does the simulation match reality?

Partly, and the parts that don't match are the point.

  • The dates line up. With the area set between about 2 and 12 m², the simulated re-entries fall between 5 and 13 February, the same window as the catalogue's 38 dates. Satellites in different attitudes would spread out just like that.
  • The standard model barely saw the storm. Fed the real storm index, the empirical model added about to the density at 210 km. SpaceX's own GPS showed up to 50% more drag. Fang et al. (2022), with SpaceX Starlink staff among the authors, ran a physics model and found at least 50% in most regions at 210 km, and more than 100% in some.
  • That gap is the problem. An operator planning with the standard model would have expected a storm worth a few hours. The real air took days. Nobody measures the density at 210 km directly; the number arrives later, worked out from how fast satellites fell.

The uncertainty view: the Sun moves the air more than anything you control

How many times denser at solar max than at solar min

Ratio of orbit-average density, F10.7 200/Ap 15 over F10.7 70/Ap 4. At your altitude: . The shaded band is ±30%, the size of the typical disagreement between good density models.

Monte Carlo lifetime uncertainty

Real lifetimes depend on things nobody knows in advance. Each run draws a random future: where we are in the 11-year solar cycle and how strong it is, the storm level, your CD (2.0–2.6), your average area (±20% for tumbling) and a ±25% model error. Then it flies the satellite down to re-entry.

Why measuring density in place matters. Every number on this page comes from a model. Models are tuned to past data. During a storm, or at a new solar minimum, they can be wrong by tens of percent just when operators need them most. In February 2022, SpaceX reported drag up to 50% higher than on earlier launches after a geomagnetic storm, and lost most of a freshly launched Starlink batch. A sensor that measures the actual density along the orbit turns these guesses into data: tighter conjunction screening, better re-entry predictions and less propellant held in reserve.

How long satellites last, and how fast a prediction drifts

Lifetime vs starting altitude · your spacecraft, Sun held fixed

Solar minModerateSolar maxYour altitude

Prediction drift from a density error

³⁄₂·Δa·t² (with the orbit's response)½·Δa·t² (simple kinematics)

Δa is the drag the model misses: the error times the drag on your spacecraft. Extra drag lowers the orbit, and a lower orbit moves faster, so the satellite runs ahead of its prediction three times faster than simple kinematics suggests (the secular along-track term of the Hill/Clohessy-Wiltshire equations). Across 151 storms, MSIS models underestimated density by about 20 to 30% (Wang et al. 2026). Both lines assume the error stays constant and nobody re-tracks the satellite.

The physics, in plain words

1. The drag equation

A satellite in low orbit moves at about 7.7 km/s through very thin air. Each second it sweeps up the air in a tube the size of its cross-section. Pushing that air out of the way costs momentum:

F = ½ · ρ · v² · CD · A

ρ is the air density, v the speed, A the area facing the flow, and CD a number near 2 that captures how molecules bounce off the surface. Divide by mass to get the deceleration. The speed is fixed by the orbit and the shape by your design, so the density is the term that swings.

2. Losing height

Drag takes energy out of the orbit, and the orbit shrinks a little each revolution. For a circular orbit of radius a:

da/dt = − (CDA/m) · ρ · √(μ a)

Lower orbits mean thicker air and faster decay, so the fall speeds up. The lifetime is the time to integrate that equation down to about 150 km, where the satellite burns up within hours. A heavy, compact satellite (high m/CDA) lasts longer.

3. Why density swings with the Sun

Above about 120 km the air is heated mostly by the Sun's extreme ultraviolet light. We track that with the 10.7 cm radio flux (F10.7), which follows the 11-year cycle from about 70 to above 200. A hotter thermosphere puffs up, so at a fixed altitude you sit in denser air: roughly 10 times denser at 400 km at solar max, and up to 30 times at 700 km.

Geomagnetic storms (high Ap) dump extra energy at high latitudes within hours. The day side, around 2 pm local time, is hotter than the night side. Switch to "one point" above to see the day/night bulge.

4. The model behind this page

Each gas (N₂, O, O₂, Ar, He) settles in its own layer, with lighter gases thinning out more slowly. That's why atomic oxygen dominates at 400 km and helium takes over higher up at solar minimum. We solve that layering in closed form (Bates' temperature profile and diffusive equilibrium) from a 120 km starting point taken from the U.S. Standard Atmosphere. The top-of-atmosphere temperature follows Jacchia's published formulas for F10.7, Ap, latitude and local time. It's about 100 lines of JavaScript (model.js).

How the simulation was built, and cross-checked

  1. Start from the requirement, not the tool. What does an operator need to know, and how well? That sets the error budget.
  2. Model the physics. Density, drag, flow into a sensor, using a reference atmosphere as the "truth".
  3. Monte Carlo the unknowns. Instead of one best guess, run hundreds of randomized cases (solar cycle, storms, instrument errors, attitude) and look at the spread. The histogram above is the same idea.
  4. Build an error budget. List each error source, size it, combine them (root-sum-square) and see which one dominates. You fix the biggest one first.
  5. Tag every number by evidence. Computed, simulated, from a datasheet, or measured. A simulated number never gets to pose as a measured one.
  6. Cross-check with an independent model. Two different implementations have to agree before a claim ships.

An AI engineering team did the heavy lifting

We developed this work during our entry to NASA TechLeap's Orbital Clarity Challenge. Two independent AI agent teams, one built on Claude and one on Grok, worked from the same requirements under one human engineer. Each team had roles like a small company: systems engineer, space-environment analyst, hardware lead, proposal writer and an independent judge that red-teams the others.

The two teams built separate simulations and then checked them against each other. Where they disagreed, we dug in, and that's how overclaims got caught before they shipped. The human set the direction and made the calls. The agents wrote the code, ran the cases, checked sources and argued.

Who built this page. Both teams, on purpose. The 3D scene and camera views started in the Claude team's viewer. The Grok team built the open density model, the drag and lifetime physics, the Monte Carlo and this simulator around it. The Claude team added the February 2022 replay (the open model driven by the real storm data, flown on a full orbit and checked against the real re-entry dates), the lifetime-vs-altitude chart and the prediction-drift chart. Goncalo Esteves directed both and asked for the two pages to become one.

This page is an example. An agent rebuilt the density model from published physics, checked it against reference models, and wrote and tested this simulator in a single afternoon. Not long ago this would have been a semester-long student project.

Satellite operators: use this, and talk to us

Use it for quick lifetime checks against the FCC 5-year post-mission disposal rule, for drag sanity checks in mission design, or to show your team why the solar cycle matters. It's free, and everything runs in the browser. If you fly in low orbit and care about density (conjunctions, re-entry, station-keeping), we'd like to hear what you need measured.

Email Goncalo: goncalo@goncaloesteves.io

Honest limits