Static copy of the whitepaper. The interactive version with the 3-D model is at colure.space/kelvinzero; the PDF is here.
How a satellite can stay out of Earth’s shadow all year — why the usual dawn–dusk orbit falls just short, at what altitude it stops falling short, and how much margin you have
A satellite in low Earth orbit normally passes through Earth’s shadow on every revolution. A dawn–dusk sun-synchronous orbit largely avoids it: J2 precession keeps the orbit plane facing the Sun all year — but it follows only the Sun’s average position, the mean Sun. In reality the Sun wanders up to 23° around this average, enough to push the plane into shadow for part of the year. We introduce MSSI, a coordinate frame that rotates with the mean Sun. In it, a sun-synchronous orbit is a fixed ring and the real Sun traces the analemma, so the yearly eclipse question becomes static geometry. From that picture we derive how much of each orbit is in shadow, the yearly eclipse for every plane orientation, maps of which directions are never, sometimes or always eclipsed, and the margin a mission has. Fixed planes become eclipse-free above about 1400 km, within a window of node offsets that widens with altitude, and the 06:00 and 18:00 lines turn out to be equivalent. The closed form is checked against a full-force-model numerical propagation and against the published eclipse seasons of three flying missions; the model, its code and an interactive 3-D visualizer are published as open source.
A satellite in low Earth orbit circles the planet every 90 to 110 minutes and, on most orbits, passes through Earth’s shadow on every lap, for up to about 35 minutes. Each pass is a small crisis the spacecraft has been designed around: the solar arrays stop producing power, so batteries must carry every load; the temperature of exposed hardware drops by tens of degrees within minutes, so heaters switch on; and when the Sun returns everything warms again. Over a five-year mission that is roughly 25,000 hot–cold cycles. Batteries, solder joints, optics and mechanisms all age with cycling, and a large share of a small satellite’s mass is there only to survive the dark half of every orbit.
An orbit that never entered the shadow would remove much of that. Power would be continuous, so the battery could be sized as a buffer rather than for every orbit’s night; temperatures would settle toward a steady state, reducing heater power and simplifying the thermal design; and payloads that depend on thermal stability — telescopes, lasers, precision clocks — would get a steadier environment. Project Kelvin Zero is an effort to find out how close a fixed orbit can get to that: can it be eclipse-free for a whole year, at what altitude, and how precisely does it have to be placed?
This paper answers that question with a small amount of orbital mechanics and one idea — a reference frame that rotates with the average Sun — that makes the geometry easy to see. The companion visualizer at colure.space/kelvinzero draws every picture in this paper live; buttons throughout the text open the matching view.
If Earth were a perfect sphere — and if the Sun and Moon did not tug on the satellite, which at these altitudesG1 they do only slightly — an orbit plane would stay fixed in space, pointing at the same distant stars. However, Earth is not a perfect sphere: it spins, and its rotation flings the equator outward so the equatorial radius is about 21 km larger than the polar radius. That extra ring of mass pulls on a satellite a little more when it is near the equator than when it is over the poles. For an orbit that is tilted with respect to the equator, the net effect of that uneven tug, averaged over many revolutions, is not to change the tilt but to slowly pivot the whole orbit plane around Earth’s spin axis — exactly as a leaning spinning top pivots around the vertical. Astrodynamicists call the bulge’s gravitational term and the pivoting nodal precession [5].
How fast the plane pivots depends on two things the mission designer chooses: how high the orbit is and how steeply it is inclined to the equator. For a circular orbit of radius (Earth’s radius km plus the altitude ) the rate is
where is the satellite’s angular speed around Earth, its inclinationG2, the size of the bulge and Earth’s gravitational constant. The sign matters: an orbit tilted less than 90° pivots westward, and an orbit tilted past 90° — one that travels slightly against Earth’s rotation, called retrograde — pivots eastward.
Seen from Earth, the Sun also drifts eastward against the stars, once around per year: per day. If the orbit plane pivots eastward at that same rate, the angle between the plane and the Sun never changes. The Sun rises over the same ground at the same local time on every pass, which is why Earth-observation satellites love these orbits — and why a plane set to face the Sun keeps facing it. Such an orbit is called sun-synchronous. Setting in (1) equal to and solving for the inclination gives
For every altitude there is exactly one inclination that works (Figure 1a, Table 1): about 97.6° at 550 km, 98.6° at 800 km, 99.5° at 1000 km. Altitude and inclination are therefore not independent choices; pick the altitude and the inclination follows. That leaves the designer with two free variables — the altitude and the orientation of the plane around the spin axis — plus the trivial choice of where along the ring the satellite starts.

| h (km) | iSSO (°) | T (min) | |β| for no eclipse (°) | α (°) |
|---|---|---|---|---|
| 300 | 96.672 | 90.52 | 72.76 | 107.24 |
| 400 | 97.030 | 92.56 | 70.22 | 109.78 |
| 500 | 97.402 | 94.62 | 68.02 | 111.98 |
| 550 | 97.593 | 95.65 | 67.02 | 112.98 |
| 600 | 97.788 | 96.69 | 66.07 | 113.93 |
| 700 | 98.188 | 98.77 | 64.30 | 115.70 |
| 800 | 98.603 | 100.87 | 62.69 | 117.31 |
| 1000 | 99.479 | 105.12 | 59.82 | 120.18 |
| 1200 | 100.419 | 109.42 | 57.31 | 122.69 |
| 1500 | 101.957 | 115.98 | 54.06 | 125.94 |
| 2000 | 104.890 | 127.20 | 49.58 | 130.42 |
The orientation of an orbit plane is described by where it crosses the equator going north, the ascending node, and the angle from a reference direction to that node is the right ascension of the ascending nodeG6, . For a sun-synchronousG5 orbit the natural reference is the Sun itself, because the node keeps a fixed angle from it. We call that angle the node offset (the slider labelled “RAAN offset from mean SunG11” in the visualizer) and it is equivalent to a clock reading, the local time of the ascending nodeG8:
puts the node under the noon Sun: the plane contains the Sun direction, the satellite flies over the sunlit face and the dark face on every lap and is eclipsed every orbit. or (LTAN 18:00 or 06:00) turns the plane to face the Sun squarely: its normal points at the Sun, sunlight falls perpendicular to the plane, and the satellite flies along the day–night line, always over dawn or dusk. This is the dawn–dusk orbit, and it is the obvious candidate for an eclipse-free mission. Figure 2 shows the three quantities together.
Interactive: Dawn–dusk plane, 800 km · Noon–midnight plane, 550 km
Equation (2) matches the orbit to the Sun’s average motion. The real Sun does not move uniformly across our sky, for two separate reasons.
Earth’s orbit is an ellipse. Earth is closest to the Sun in early January (perihelionG28) and farthest in early July (aphelion). By Kepler’s second law it moves fastest when closest, so the Sun’s apparent eastward drift is about 7% faster in January than in July. Over a season the Sun runs ahead of, then behind, its average position by up to about 2°.
Earth’s axis is tilted. The spin axis leans 23.44° from the perpendicular to the orbit plane. The orbit plane of a satellite pivots around that tilted axis, but the Sun moves along the ecliptic, the plane of Earth’s orbit. Projected onto the equator — the only direction the satellite’s plane can follow — the Sun’s eastward motion speeds up near the solstices and slows near the equinoxes, and the Sun also swings 23.44° north and south of the equator over the year. The north–south swing is the seasons; the east–west speed-up is a second contribution to the Sun running ahead or behind.
Because of these two effects, astronomers long ago defined a convenient fiction called the mean Sun: a point that moves eastward along the celestial equator at the perfectly uniform rate of per day. It is the Sun we would have if Earth’s orbit were circular and its axis upright. Mean solar time — the time on every clock — is defined by it. A sun-synchronousG5 orbit is, precisely, an orbit locked to the mean Sun: its plane follows the mean SunG11 flawlessly and the real Sun wanders around that mean position. The size of the wander is the whole story of this paper: about east–west (the famous equation of time, minutes on a sundial) and north–south.
A dawn–duskG9 orbit therefore faces the mean Sun squarely — its normal points at the mean Sun — and the real Sun shines on it from up to 23° above or below that plane. Whether that is enough to push the satellite into shadow depends on the altitudeG1 — a higher orbit looks over Earth’s shoulder more easily — and on how far the node sits from the ideal 06:00 or 18:00 line. The rest of the paper quantifies exactly that.
To reason about geometry that changes over a year we need a coordinate frame, and the right choice makes the problem almost trivial.
ECI — Earth-Centred Inertial. The standard inertial frame of orbital mechanics, used here in its J2000G19 realisation (the mean equator and equinox of J2000, “EME2000”; the GCRF differs from it by a few hundredths of an arc-second, which is irrelevant at this level). Its origin rides along with Earth around the Sun, but its axes never rotate: points along Earth’s spin axis and toward a fixed direction among the stars, the vernal equinoxG16 — the direction of the Sun on 20 March. In ECIG17 a sun-synchronousG5 orbit plane is seen to pivot once per year, and the Sun circles once per year along the eclipticG15.
MSSI — Mean-Sun Synchronous Inertial. The frame we introduce for this problem. It is ECI rotated about the spin axis so that its axis always points at the mean Sun. If is the mean SunG11’s longitude measured from the vernal equinox on day (days from the J2000 epoch, 2000-01-01 12:00),
The frame turns once per year — exactly as fast as a sun-synchronous plane precesses. The payoff is that the precessionG4 is baked into the frame: in MSSIG18 a sun-synchronous orbit is one fixed circle around Earth that never moves, the mean Sun is a fixed point on the axis, and the only thing that moves over the year is the real Sun, wandering around that point. The yearly eclipse problem becomes a static picture: a fixed ring, and a small loop that the Sun traces beside it. Figure 3 shows both frames carried around the Sun, and Figure 4 and Figure 5 show the Sun-centred view in the visualizer. The visualizer is the best place to see the difference: the two views linked below show the same orbit with the frame planes drawn, once in ECI — press play and the plane visibly pivots as the year passes — and once in MSSI, where it stays put while the Sun wanders.
Earth’s daily rotation is independent of either frame: Greenwich sits at sidereal angle from the equinox, so the globe is turned by in MSSI. At (J2000, 12:00 UT) that is 0.0006°: Greenwich lies under the mean Sun at noon, which is one of the sanity checks used by the end-to-end tests.




Interactive: Orbit and frame planes in ECI — the plane pivots · The same in MSSI — the plane stays put · Sun-centred view (Earth orbiting the Sun)
What does the real Sun do in the mean-Sun frame? Its position needs only a short ephemeris — the U.S. Naval Observatory’s low-precision solar formulae [1], accurate to about one arc-minute over 1950–2050 — which we use for every Sun-dependent quantity in the model so that the sub-solar pointG20, the Sun vector, the terminatorG21, the eclipse tests and the frame rotation are consistent by construction:
Here is the mean anomaly (where Earth would be on a circular orbit), the terms are the equation of centre — the correction for the elliptical orbit — is the Sun’s true longitude along the eclipticG15, the axial tilt, and are the Sun’s right ascension and declinationG13: its longitude and latitude on the celestial equator. The difference between the real and mean SunG11 along the equator is the equation of time, , positive when a sundial runs ahead of the clock [2]. In MSSIG18 the Sun therefore sits at longitude and latitude .
Plot that point over a year and you get a slender figure-8 (Figure 6): the analemma. It is the same shape that appears in a photograph of the Sun taken from one spot at the same clock time every week — a picture first captured on film by Dennis di Cicco in 1978–79 — and the same curve printed on old globes and sundials to convert sundial time to clock time. The equation of timeG12 itself has been known since antiquity: Ptolemy tabulated it in the Almagest in the 2nd century, and it was the great nuisance of navigation and timekeeping until mechanical clocks replaced the Sun as the standard. The two lobes of the 8 are unequal because the two causes — the elliptical orbit (a once-a-year wave) and the tilt (a twice-a-year wave) — add with different periods; the bigger lobe lies in the southern, northern-winter half of the year, when both effects pull the same way.

There is a simpler way to see why the figure-8 has to appear here. The classic analemmaG14 photograph is made by photographing the Sun from the same spot at the same clock time every week or so. Clock time is mean solar time, so “the same clock time” means “when the mean Sun is in the same place in the sky”: the photographer is recording where the real Sun is relative to the mean Sun. That is exactly what the MSSI frame does continuously — it holds the mean Sun still and records the real Sun — so the photographer’s figure-8 (Figure 7) and ours are the same curve, drawn from Earth’s centre instead of from a roadside. The connection to our problem is now direct. The analemma is the set of directions from which sunlight arrives over the year, drawn in the frame where the orbit is frozen. Its height, , is the Sun’s excursion above and below the plane of a dawn–duskG9 orbit; its width, , is how far the Sun strays east–west from the direction the plane faces. Figure 8 shows the curve on the globe, with the Sun on it, in the visualizer.


Interactive: Analemma at the December solstice · Analemma at the June solstice
With the Sun effectively at infinity, Earth casts a cylindrical shadowG23 of radius pointing directly away from the Sun. A point at position is in shadow when it is on the night side and within one Earth radius of the shadow’s axis:
where is the unit vector toward the Sun. For a circular orbit the only number that matters is the β angle: the elevation of the Sun above the orbit plane, with the plane’s normal (Figure 9). β lives in the plane spanned by the orbit normal and the Sun direction: turn the picture until you look along the orbit plane with the Sun still in view, and the ring collapses to a line with the Sun β above it, which is how Figure 9b is drawn. When the Sun lies in the plane and the satellite crosses the full width of the shadow; as grows the ring is tilted out of the shadow and the dark arc shrinks, until the whole ring clears the cylinder. Writing , the fraction of each revolution spent in shadow is
The orbit is eclipse-free once — 67.0° at 550 km, 62.7° at 800 km, 59.8° at 1000 km (Figure 1b). This is where altitudeG1 enters: a higher satellite sees over Earth’s edge at a shallower angle, so it tolerates a Sun that is further from the ideal direction. Figure 10 plots (8); Figure 11 shows the shadow test applied to a noon–midnightG10 ring in the visualizer.


For a dawn–duskG9 orbit the normal points almost at the mean SunG11 — not exactly, because a sun-synchronousG5 orbit is tilted 8° or so past the pole, so the normal sits at latitude . The β angleG22 is then 90° minus the angle between that normal and the real Sun, which moves around the analemmaG14. Near one solstice the Sun is 23° on the far side of the normal and β drops to about 58°; near the other it is on the near side and β is nearly 90°. Whether 58° is enough depends on the altitude: at 800 km it is below the 62.7° threshold, so the orbit has a short eclipse seasonG26 around that solstice; at 1500 km the threshold is 54° and the orbit never eclipses. The yearly exposure is simply the average of over the days of the year.
Interactive: Eclipse segments from the anti-Sun side
The β-angleG22 argument answers the question one orbit at a time. The mean-Sun frame allows something more general: a map, drawn once per altitudeG1, that shows at a glance which orbits are eclipse-free. Think of the sphere of all directions around Earth, seen from the centre, in MSSIG18. A satellite at altitude lying in direction is sunlit exactly when the angle between and the Sun is less than a cone half-angle that depends only on the altitude:
Because the Sun’s direction is a fixed curve in MSSI — the analemmaG14 — the whole year’s behaviour of every direction is captured by two numbers, the largest and smallest value of over the year:
Each direction is coloured by what a point of the orbit shell there sees over the year: green — in sunlight all year (inside ); orange — in sunlight for part of the year and in shadow for the rest (inside but outside ); purple — in shadow all year (outside ). The boundaries and are contours of those two numbers. An orbit at that altitude is a great circle on this sphere, so the map answers the eclipse question at a glance: (1) if the whole ring lies in green, the satellite is never eclipsed at any time of the year; (2) if the ring passes through orange but not purple, it has an eclipse season — eclipse-free for part of the year, an eclipse on every revolution for the rest; (3) if the ring crosses purple, it is eclipsed on every revolution all year round. Note that orange does not mean “a little shadow on every orbit”: on a given day a point is either sunlit or not, so an orange crossing means whole seasons with eclipses and whole seasons without. Figure 12 shows the maps for 550 km and 1000 km; Figure 13 drapes one on the globe with a noon–midnightG10 ring crossing the purple cap. Raising the altitude shrinks the purple and orange regions, which is the map’s way of saying what (8) said: higher orbits tolerate more.


Interactive: Eclipse-time regions · Regions rotating in ECI
Figure 14a sweeps the node offsetG7 through a full turn for five altitudesG1. The curve has two deep valleys at the dawn–duskG9 orientations and two broad hills at noon–midnightG10, where the plane contains the Sun and the satellite spends 30–37% of the year in shadow. The valleys are not identical. With the orbit normal sitting 8° south of the mean SunG11, (LTANG8 06:00) points the sunward normal north, so its eclipse seasonG26 falls around the December solstice; because perihelionG28 is in early January the Sun spends about a week less at southern declinations, and in the closed form 06:00 edges out 18:00: 4.83% versus 5.03% of the year at 550 km, 2.67% versus 2.78% at 800 km. The difference is small — ten hours a year at 800 km — and Section 10 shows that the full force model takes it back: the higher zonal harmonics of the gravity field, which the closed form omits — chiefly , which gives the orbit a small forced eccentricity whose orientation relative to the Sun differs between the two lobes — are worth about the same ten hours in the other direction, so the two lobes come out within two hours a year of each other. For eclipse purposes the two dawn–dusk lines are interchangeable; the choice between them can be made on other grounds, such as the lighting of the ground track or ground-station timing. Figure 14b shows the matching β seasons: the 1500 km plane never crosses its threshold.

At the best orientation the number of days per year with any eclipse falls from about 97 at 550 km to 74 at 800 km, 56 at 1000 km and 36 at 1200 km (Figure 15). From roughly 1400 km upward a fixed plane exists whose β never drops below the threshold — a truly eclipse-free sun-synchronousG5 orbit without any node control. Below that altitude reaching zero would require the plane to follow the Sun’s declinationG13 through the year — an “analemmaG14-following” node schedule, maintained with propulsion — which is outside the scope of a fixed plane.

Figure 16 puts every fixed sun-synchronous plane on one chart: altitude against node offset, coloured by yearly eclipse, with the eclipse-free island outlined in green. Figure 17 reads the margins off that chart: how many degrees the node may sit away from the ideal dawn–dusk line while staying under a given yearly eclipse. The eclipse-free window opens at 1400 km only 3.7° wide (node offsets 267.7°–271.4°), grows to 12.4° at 1500 km (263.4°–275.8°), 16.8° at 1600 km and 25.6° at 2000 km (256.8°–282.4°); it sits almost symmetrically about the 270° line, a few tenths of a degree low. If a mission can accept 1% of the year in shadow (about 90 hours) a window exists from about 1110 km, 10° wide at 1150 km and 14° at 1200 km; at 2% it opens near 900 km, and at 5% — still only 440 hours a year — a 600 km orbit has a 16°-wide window (about ±8° around the dawn–dusk line) and an 800 km orbit 32° (about ±16°). In local time, 15° is one hour of LTAN. Figure 18 shows the same map as a relief. Table 2 collects these design points.





Figure 19 shows two dawn–dusk planes on their worst day of the year against the eclipse-time regionsG27 of their shells. In the units a power or thermal engineer works in, the longest eclipse of the year at the best orientation is 21.7 minutes per 95.6-minute orbit at 550 km, 16.9 minutes at 800 km, 13.0 at 1000 km and 8.7 at 1200 km — against about 35 minutes every orbit for a noon–midnight plane — and the yearly totals are 424, 234, 130 and 54 hours. From 1400 km the longest pass is zero. Table 2 lists these numbers with the inclinationG2, the β margin and the node window for each altitude.
| h (km) | i (°) | T (min) | min |β| (°) | threshold (°) | eclipse (h/yr) | days | longest pass (min) | eclipse-free Ω (°) | width (°) | ≤ 1 % window (°) |
|---|---|---|---|---|---|---|---|---|---|---|
| 550 | 97.59 | 95.6 | 59.0 | 67.0 | 424 | 97 | 21.7 | — | — | — |
| 800 | 98.60 | 100.9 | 58.0 | 62.7 | 234 | 74 | 16.9 | — | — | — |
| 1000 | 99.48 | 105.1 | 57.1 | 59.8 | 130 | 56 | 13.0 | — | — | — |
| 1200 | 100.42 | 109.4 | 56.1 | 57.3 | 54 | 36 | 8.7 | — | — | 14.0 |
| 1400 | 101.43 | 113.8 | 55.1 | 55.1 | 0 | 0 | 0 | 267.7 – 271.4 | 3.7 | 24.0 |
| 1500 | 101.96 | 116.0 | 54.6 | 54.1 | 0 | 0 | 0 | 263.4 – 275.8 | 12.4 | 27.2 |
| 1600 | 102.51 | 118.2 | 54.1 | 53.1 | 0 | 0 | 0 | 261.2 – 278.0 | 16.8 | 29.8 |
| 1800 | 103.66 | 122.7 | 52.9 | 51.3 | 0 | 0 | 0 | 258.4 – 280.7 | 22.3 | 34.0 |
| 2000 | 104.89 | 127.2 | 51.7 | 49.6 | 0 | 0 | 0 | 256.8 – 282.4 | 25.6 | 36.9 |
Interactive: Kelvin Zero plane, 1500 km dawn–dusk · 800 km dawn–dusk in its eclipse season
Everything above reduces to a short recipe that needs no simulation. Given an altitudeG1 and a node offsetG7 :
The visualizer’s sidebar performs this calculation live for the active orbit — period, LTANG8, the current β and shadow fraction, the eclipse-free threshold, the β range and sun exposure over the year, and the number of days with eclipse — and draws β over the year with the eclipse band shaded. The model is deliberately simple: a circular orbit, J2G3 secular precessionG4 only, a cylindrical shadowG23 and the Sun at infinity. Section 11 lists what that leaves out and Section 10 measures it: against a year of numerical propagation with the full gravity field, the Sun and Moon and a conical shadow, the closed form is high by 8–23 hours a year in umbraG24 time (0.1–0.26% of the year) and within about 8 hours once the penumbra is counted, and neither the season dates nor the eclipse-free threshold move.
Two independent implementations. The model behind this paper is implemented twice: once in JavaScript, which drives the visualizer and every figure here, and once in Python, written separately from the same equations. The Python implementation produces reference values — the solar ephemeris, the sun-synchronousG5 inclinationG2 and precessionG4 rate, Kepler-solved and series analemmas, brute-force cylindrical-shadow sweeps over sampled orbits, and the eclipse-region extrema — and an automated test suite compares the JavaScript model against them on every change. The two agree on the solar ephemeris and the sun-synchronous inclination to degrees, on yearly eclipse fractionG25s to within 0.0002% of the year — about a minute — (closed form (8) against brute-force sampling of the cylindrical shadowG23 at 7200 points per orbit on 3660 days), on the eclipse-region extrema to , and on the analemmaG14 extents to 0.06°. Orientation conventions are tested explicitly: the Sun is west of the mean SunG11 in November, an exact sun-synchronous plane is frozen in MSSIG18, and gives LTANG8 18:00.
Scene against model. An end-to-end harness drives the visualizer’s interface and reads world-space positions back from the rendered scene: the sub-solar marker lies on the analemma on every sampled day in both frames; the ring’s vertices move by less than Earth radii over a year in MSSI and its node advances 89.70° in 91 days in ECIG17 (expected 89.69°); the satellite stays at radius ; the globe’s rotation is 0.0006° at and 180.0° at , i.e. Greenwich is under the mean Sun at J2000G19 noon.
Against the literature. The USNO low-precision solar formulae are compared with the widely used NOAA/Spencer compact series [4] over a full year: they differ by at most 0.9 minutes in the equation of timeG12 and 0.5° in declinationG13, both well below anything that affects the conclusions. The closed-form eclipse fraction (8) is the standard result for a circular orbit in a cylindrical shadow [6]; it is worth noting that a tempting simplification, with threshold , is wrong — it predicts no eclipse beyond at 550 km, where the geometric threshold is 67° — and (8) is confirmed against the brute-force sweeps to within 0.0002% of the year.
The closed form is only as good as its three simplifications — J2G3 secular precession, a cylindrical shadow, a one-arc-minute solar ephemeris — so we propagated nine orbits numerically through a full year with none of them: Earth’s zonal gravity field through (EGM96), the Sun and Moon as point masses with positions from the JPL DE440 ephemeris, DOP853 integration at a relative tolerance of , and every 10 s a conical umbraG24/penumbra test with the real Sun–Earth distance and solar disc (pass edges refined to a millisecond). Each case is run twice: with the inclination of equation (2), which measures how far the J2-only sun-synchronous condition is from the truth, and with that inclination corrected once so that the node is stationary relative to the mean Sun under the full force model — what a mission would fly. The whole procedure is kelvinzero hifi in the open package and takes a few minutes per case. Figure 20 overlays the daily eclipse minutes of the two models for the 800 km and 1200 km dawn–duskG9 planes; Table 3 collects the year totals.

| h (km) | LTAN | closed form (h/yr) | days | longest (min) | hi-fi umbra (h/yr) | days | longest (min) | penumbra (h/yr) | drift (°/yr) | Δi (°) |
|---|---|---|---|---|---|---|---|---|---|---|
| 550 | 12:00 | 3261.8 | 365 | 35.6 | 3245.2 | 365 | 35.5 | 26.0 | -0.62 | +0.0132 |
| 550 | 06:00 | 423.8 | 97 | 21.7 | 411.2 | 97 | 21.4 | 31.4 | -0.45 | +0.0095 |
| 800 | 06:00 | 233.9 | 74 | 16.9 | 222.5 | 74 | 16.3 | 28.6 | -0.42 | +0.0102 |
| 800 | 18:00 | 243.5 | 77 | 16.9 | 220.6 | 73 | 16.0 | 29.0 | -0.48 | +0.0114 |
| 1000 | 06:00 | 129.8 | 56 | 13.0 | 122.0 | 55 | 12.7 | 26.4 | -0.41 | +0.0108 |
| 1200 | 06:00 | 53.6 | 37 | 8.7 | 44.2 | 34 | 8.1 | 24.5 | -0.38 | +0.0111 |
| 1400 | 06:00 | 0.0 | 0 | 0.0 | 0.0 | 0 | 0.0 | 9.8 | -0.35 | +0.0113 |
| 1500 | 06:00 | 0.0 | 0 | 0.0 | 0.0 | 0 | 0.0 | 0.0 | -0.34 | +0.0114 |
| 1600 | 04:52 | 118.1 | 58 | 12.9 | 108.1 | 57 | 12.4 | 24.7 | 0.11 | -0.0039 |
Three things stand out. First, the closed form is accurate. The propagation’s umbra hours are 8–23 hours a year below it (0.1–0.26% of the year), the days with eclipse agree to within three days, the longest pass to within a minute, and the eclipse-free cases stay eclipse-free: at 1400 km, the edge of the island, the satellite spends 9.8 hours a year in partial shadow, losing the equivalent of 1.6 hours of full sunlight, and never enters the umbra. The umbra deficit is just the cone — 8 to 10 s per pass times a thousand passes — and if the penumbra is counted by the fraction of the Sun it hides, which is what a solar array sees, the two models agree to within about 8 hours a year, because the cylinder is in effect the umbra plus half the penumbra. Second, the J2-only inclination of equation (2) is within 0.013° of the right value; left uncorrected it lets the node drift by 0.3–0.6° a year on the dawn–dusk cases (0.1° on the 1600 km case away from the line), which is negligible against the 12–26° windows above 1500 km but is a decade’s worth of the 3.7° window at 1400 km. Third, the small advantage of 06:00 over 18:00 in Section 8 does not survive: at 800 km the propagation finds the two lobes within two hours a year of each other (222.5 against 220.6 h), so the perihelionG28 effect is real but cancelled. Re-running the 800 km pair with J2 alone restores the closed form’s 9.5-hour advantage for 06:00; adding – removes it (221.1 against 220.3 h), and the Sun and Moon then move the totals by only a couple of hours. The culprit is the pear-shaped term, which forces a small eccentricity on a sun-synchronous orbit whose orientation relative to the Sun differs between the two lobes.
Three solar observatories fly dawn–dusk sun-synchronous orbits for exactly the reason this paper studies, and their operators publish when the eclipses come. Table 4 sets those published seasons beside the model’s prediction from nothing but the altitudeG1 and the side of the dawn–dusk line. The dates agree to within a few days and the longest pass to within about a minute — including the fact that the two 06:00 missions have their seasons around the December solstice and Hinode, on the 18:00 side, around the June solstice, which is the asymmetry of Section 8 seen in flight. The published seasons are for real, slightly elliptical orbits with their own node drift, so this is as close as a closed form can be expected to get.
| Mission | Orbit | Published eclipse season | This model |
|---|---|---|---|
| PROBA-2 (ESA, 2009–) | 725 km, i = 98.3°, LTAN 06:00 | about 80 days, November to January, eclipses up to 18 min per orbit | 11 Nov – 31 Jan, 80 days, up to 18.3 min |
| IRIS (NASA, 2013–) | 620 × 670 km, i = 97.9°, LTAN 06:00 | eclipse-free from the start of February to the end of October; 7–8 months of continuous observation | (645 km) 7 Nov – 3 Feb, 88 days, up to 19.8 min |
| Hinode (JAXA/NASA, 2006–) | 680 km, i = 98.1°, dawn–dusk, 18:00 side | early May to early August, nights up to 20 min per 98-min orbit; nine months of continuous sunlight | 8 May – 4 Aug, 87 days, up to 19.2 min |
Since late 2025 three organisations have published plans for large computing platforms in sun-synchronous orbit, chosen for exactly the continuous sunlight this paper is about. Google’s Project Suncatcher describes an 81-satellite cluster in a 650 km dawn–dusk orbit, with two prototypes planned for 2027 [12]. Starcloud’s FCC application places up to 88,000 satellites in shells of up to 50 km between 600 and 850 km, sun-synchronous with a nominal 06:00 crossing time [13]. SpaceX’s January 2026 FCC application for an “Orbital Data Center System” covers up to one million satellites in 50 km shells from 500 to 2000 km, in sun-synchronous and 30° inclinations, with representative sun-synchronous planes at 500, 1000 and 2000 km in its Schedule S — the altitudes used in Table 5 [14]. Starcloud’s filing gives a nominal 06:00 crossing time and Google’s paper a dawn–dusk orbit; SpaceX’s filing gives no local time, and for its sun-synchronous shells we assume the dawn–dusk line, the only choice consistent with its stated purpose; where a paper or filing names no side, the 06:00 line of Table 2 is used. Table 5 runs each through the closed form, and every one of these orbits is a preset in the visualizer.
| System | Filed orbit | Eclipse (h/yr) | Days with eclipse | Longest pass (min) | Sunlit (%) |
|---|---|---|---|---|---|
| Google Project Suncatcher | 650 km, dawn–dusk SSO | 338 | 87 | 19.7 | 96.2 |
| Starcloud | 600–850 km, SSO shells, 06:00 crossing (as filed) | 379 (600 km) to 204 (850 km) | 92 to 68 | 20.7 to 16.0 | 95.7 to 97.7 |
| SpaceX Orbital Data Center System | 500–2000 km, SSO shells | 474 (500 km), 130 (1000 km), 0 from 1400 km | 102, 55, 0 | 22.7, 13.0, 0 | 94.6, 98.5, 100 |
| SpaceX Orbital Data Center System | 500–2000 km, 30° shells (not sun-synchronous) | 3160 (500 km) to about 2000 (2000 km) | 365 (342–355 at 2000 km) | 35.8 to 35.0 | 64 to 77 |
The pattern is the paper’s result seen from the other side. Every announced sun-synchronous shell below 1400 km has an eclipse seasonG26 of two to three months, with passes of 13 to 23 minutes; Suncatcher at 650 km would spend about 340 hours a year in shadow, the same regime as the solar observatories in Table 4. Only the top of SpaceX’s range reaches the eclipse-free island, and at 1400 km the window is 3.7° wide. SpaceX’s narrative makes the case in the same terms as this paper — sunlight “up to more than 99% of the time for high-altitude, sun-synchronous orbits”, with the lower-inclination shells load-balancing compute against time-variable demand [14]; the closed form puts that 99% line at about 1110 km for a dawn–dusk plane (98.5% sunlit at 1000 km, 99.4% at 1200 km) and 100% from 1400 km. The lower altitudes are not an oversight: debris-mitigation rules, the inner radiation belt above about 1000 km, launch mass and link latency all argue for flying low, and the 30° shells trade sunlight for cheaper launch and lower latency: at 500 km they are in shadow 36% of the time, about 3200 hours a year, in passes of up to 36 minutes. What the model adds is the price of each choice in hours of darkness, battery depth and thermal cycles — the quantities a power and thermal design is sized by.
Both implementations are open source. The Python model is published as the kelvinzero package [11] (pip install kelvinzero): a few hundred lines of NumPy organised exactly as this paper is, with the test suite, a command-line tool that prints the statistics for any orbit, a script that regenerates every figure here, and a self-contained 3-D viewer, so that every number in this paper can be regenerated and inspected locally.
The shadow is cylindrical: there is no penumbraG24 and no atmospheric refraction, and the Sun is at infinity. With the real Sun–Earth distance the umbra is a cone of half-angle 0.26°, which at β = 0 shortens a pass by 8.4 s at 550 km, 8.9 s at 800 km and 10.2 s at 1500 km out of about 35 minutes, while the penumbra adds a partial fringe of 8.5–10.4 s at each end; Section 10 shows what that does to a year. Orbits are circular and propagated with the secular J2G3 rate only; Section 10 supplies the full numerical cross-check (J2–J6, the Sun and Moon, a conical shadow), which moves eclipse hours by a few percent at most, and Orekit’s propagators [7] are an independent way to repeat it. Earth is a sphere of equatorial radius. Natural extensions are the analemmaG14-following node schedule, a Sun-centred inertial frame for the dynamics, Earth-fixed ground tracks with repeat cycles, and a conical umbra/penumbra term in the closed form itself. The model layer is framework-free and unit-tested, so each can be added with a reference value in the verification script.
This paper isolates one question — can a fixed plane stay out of the shadow — and answers it. Choosing an orbit for a real mission weighs that answer against other things that the same altitudeG1 band changes, and it is worth naming them, without analysing them here:
Acknowledgements and credits. The visualizer is built with React, three.js and KaTeX (MIT licences); its Earth imagery is NASA Blue Marble, public domain, via the example textures of three-globe (MIT). The model is NumPy; the high-fidelity check uses SciPy and jplephem, reading the JPL DE440 ephemeris (NASA/JPL). The typefaces are Source Serif 4, Inter and JetBrains Mono under the SIL Open Font License. The analemma photograph is by Giuseppe Donatiello (CC0) [3]. Full third-party notices: colure.space/kelvinzero/THIRD-PARTY-NOTICES.txt.
The terms used in this paper, in the order a reader meets them. In the text, the first use of each term in a section carries a small G-number pointing here; figure references point to the picture that shows each one.
kelvinzero.hifi, reference 11).pip install kelvinzero); and the interactive visualizer at colure.space/kelvinzero.Cite as: K. Cooper and Y. Nikas, Eclipse-minimising sun-synchronous orbits in a mean-Sun rotating frame, Project Kelvin Zero technical whitepaper (October 2026), arXiv:2611.19726, colure.space/kelvinzero. Model and code: doi:10.5281/zenodo.23382386 (github.com/kylecooper7/kelvin-zero, pip install kelvinzero).
Use of AI tools: Generative AI was used throughout this work, including to draft and edit the text, to write the model, test, propagation and figure code, and to build the visualizer. The problem, the mean-Sun frame, the design of the verification and every claim are the work of the authors. The authors checked every derivation, number and reference, verified the closed form independently against a full-force propagation and three flying missions, and take full responsibility for the content of this work.
Constants: km³/s², km, , tropical year 365.2422 d, mean-Sun rate 0.9856474°/day, obliquity 23.439°, , J2000.0 = 2000-01-01 12:00.