A distance-accurate exoplanet explorer where every pixel can name its evidence.
Fly from one astronomical unit to a two-metre observer height on a real exoplanet — with correct positions, real stellar colours, and an honest label on everything that is inferred rather than measured.
The TRAPPIST-1 planets. Artist concepts — nobody has ever resolved an exoplanet surface, and that fact is the whole design problem. NASA/JPL-Caltech, public domain.
Three kinds of exoplanet software exist, and none of them do what Aether does.
| Strength | Limitation | |
|---|---|---|
| Catalogs — NASA Exoplanet Archive, ExoFOP | Authoritative | Tables. No sense of place. |
| Visualizers — OpenSpace, Gaia Sky, NASA Eyes | Correct orbits | Planets are spheres. You cannot land. |
| Space games — No Man's Sky, Elite, KSP | Immersion, agency | Entirely invented. Zero traceability. |
Aether occupies the empty quadrant: immersive and auditable at the same time.
For almost every exoplanet there is no spatially resolved surface observation. Transmission spectra constrain atmospheres, not geography. Mass and radius constrain a class of interior, not a coastline.
So Aether does not claim to reconstruct a planet's true surface. It claims this instead:
Aether generates deterministic, physically admissible surface hypotheses conditioned on measured system data, retrieval posteriors, interior models, and Solar System analog data. Every visible feature is labelled by evidence class and confidence.
The word accurate is reserved for orbits, distances and unit conversions. Generated geography is scientifically constrained, never accurate and never predicted.
What standing on TRAPPIST-1f might look like. "Might" is doing all the work in that sentence, and Aether is built to keep saying so. NASA/JPL-Caltech, public domain.
Pick a planet with three measured parameters and the hypothesis envelope is wide: the surface is a guess, and it is drawn with a strong amber speculation rim to say so. Then add evidence:
| Add | And you gain |
|---|---|
| Gaia | A stellar distance, so the transit depth becomes a real radius |
| Radial velocity | A mass, therefore density and surface gravity |
| JWST | An atmosphere, so the sky colour is derived rather than assumed |
The envelope narrows, hypotheses are eliminated, the world re-composes, and the amber rim recedes. The discovery loop, the terrain generator, the provenance system and the uncertainty machinery are not four features — they are one mechanic.
These four surfaces were produced by the terrain harness from TRAPPIST-1's published measurements. Nothing about the geography is observed — no exoplanet surface ever has been — but nothing about it is arbitrary either. A model selects operators from a fixed catalog, and physics decides which operators it is even allowed to see.
 TRAPPIST-1 b · lava_world · 398 K · 4.17 S⊕1,833 W/m² of tidal heating — five hundred times Io. Magma ocean, shield volcanoes and flood basalt. Dark rather than glowing: radiating that flux away settles the surface at 490 K, far below the basalt solidus. |
 TRAPPIST-1 e · cratered_relict · 250 K · 0.65 S⊕Bright cratered highlands cut by dark flood-basalt plains — the lunar maria pattern. The best-known habitable-zone candidate in the system. |
 TRAPPIST-1 g · ice_world · 172 K · 0.25 S⊕Ice shell, glacial carving and a tidal fracture network over a cratered base. |
 TRAPPIST-1 h · ice_world · 149 K · 0.14 S⊕The outermost planet, and the coldest. Hypsometry matched to an icy-satellite reference. |
The archetypes were not assigned by hand. They fall out of the measured numbers. All four planets orbit the same 2566 K dwarf at the same age and the same distance from us, so the only thing separating a lava world from an ice world here is where it sits: 4.17 down to 0.14 times Earth's illumination. Insolation is derived from TRAPPIST-1's own measured luminosity — 5.55 × 10⁻⁴ L☉ — because assuming a Sun-like host would overstate it a thousandfold and turn every one of these into a lava world.
What the model may and may not do. It picks operators and their parameters from a
registry. It cannot invent a process: rivers are not discouraged on an airless rock, the
operator is absent from the catalog, absent from the JSON schema, and rejected by the
validator if it appears anyway. Parameters outside a physically admissible range are clamped
and the correction is recorded. Every value carries measured, derived, assumed,
ai_inferred or repaired, so you can always ask which of those a given number was.
Specs are frozen and content-addressed, so the same planet renders the same way every time — asking twice is a ledger hit, not a second model call.
# The cards are generated from the seed data, so they are built rather than committed.
pnpm --filter @aether/catalog run export:trappist
cd tools/aether-harness
export OPENCODE_API_KEY=... # any OpenAI-compatible provider works
python build_incremental.py --cards ../../packages/catalog/trappist-cards.jsonThe Solar System runs through the identical path, and that is the point: it is the only place with ground truth, so it is where the generator can be shown to be wrong. It currently gets Mercury and the Moon as airless rock, Venus as a greenhouse at 251 K, Earth capped at 8,850 m of relief — which is Everest — Mars as a cratered relict, all four giants as volatile envelopes, and Io as a lava world.
Aether models fourteen missions. Each declares its instruments, wavelength coverage, and — the part that matters — exactly which physical parameters its data can constrain.
Selecting missions is a scientific act. Choose TESS alone and you have a radius ratio and a period; parameter coverage sits under 25% and the surface is mostly invention. Choose "everything flying today" and it climbs past 70%.
Two rules keep this honest:
- Quality is the best across missions, never the sum. Ticking more boxes cannot manufacture confidence.
- The envelope never fully closes. Even with every mission selected it bottoms out at 15%, because no exoplanet surface has ever been resolved. That floor is not pessimism.
The Habitable Worlds Observatory is the only mission that constrains a surface at all — and even then only weakly, because published work finds ocean and sand nearly indistinguishable from 0.4–1.8 µm, with a severe degeneracy between planetary radius and surface albedo.
The star field is 109,389 real stars at their measured parallax distances, from the HYG database — Hipparcos for the distances, the Yale Bright Star Catalogue for the naked-eye sky, Gliese for nearby stars. These are the same underlying catalogues OpenSpace's Digital Universe star layer is built from.
Two things make it a map rather than a dome:
- Positions are three-dimensional and real. The constellations are an accident of our viewpoint, and would distort if you travelled far enough. Nothing is painted on a sphere.
- Brightness is computed, not stored. Apparent magnitude comes from the distance
modulus,
m = M + 5·log₁₀(d) − 5, so approaching a star brightens it by exactly the right amount.
Colour runs B−V colour index → effective temperature (Ballesteros 2012)
→ blackbody → sRGB, through the same stellarColor() that lights the planets. One
physical model, applied consistently.
The tests check it against the real sky: the brightest star comes out at magnitude −1.46 (Sirius), there are ~8,700 stars brighter than the naked-eye limit against the ~9,000 usually quoted, and the count roughly triples per magnitude step as it should.
Loading is progressive — the naked-eye tier is 170 KB and appears immediately; the full 2.1 MB catalogue streams in behind it.
On licensing: OpenSpace renders from the AMNH Digital Universe atlas, whose licence requires written permission to redistribute inside software. HYG carries the same science from the same sources under CC BY-SA 4.0. Going to the shared upstream is both the more permissive route and the better-provenanced one. See
packages/stars/data/ATTRIBUTION.md.
Every marker is filled with its star's true blackbody colour, computed by the same
stellarColor() function that lights the 3D scene. TRAPPIST-1 at 2566 K really is that
orange, and a world lit by it looks nothing like a world lit by the Sun. We do not correct
for it — an M-dwarf world is genuinely dim, and that is a teaching moment, not a bug.
Every package is a building block with one responsibility and no circular dependencies. Dependencies point strictly downward, so any layer can be replaced without touching the ones beneath it.
apps/
web/ Three.js client — the only package that knows about rendering
packages/
catalog/ Typed seed data with provenance -> derived physics -> PlanetModel
exoplanets/ Every confirmed system, mapped, with priors for what was never measured
deepsky/ Galaxies, clusters and nebulae (OpenNGC) + derived distances
stars/ Real star catalogue (HYG) + physical photometry
solar-system/ Real J2000 elements, Keplerian mechanics, the terrain validation path
terrain/ Terrain operators + physics admissibility gates + program compilation
missions/ Mission registry + the evidence model
physics/ Formula registry + deterministic derivations with full provenance
schema/ The contract: Science / Inference / Visual layers, LLM containment
coordinates/ f64 nested reference frames, floating-origin camera, depth partitioning
units/ SI discipline — branded quantities, IAU/CODATA constants, conversions
web ──► solar-system ──► terrain ──► physics ──► units
│ └──────► schema ───► units
├──────► missions ─────────────────────────────► units
└──────► coordinates ──────────────────────────► units
The renderer has already changed twice — C++/OpenGL → Babylon.js → Three.js — and the
science layers needed zero rework each time, because everything above apps/ is written
against the contract rather than against an engine. That line is worth holding.
interface PlanetModel {
science: ScienceSpec; // ingestion + allowlisted physics affects science: yes
inference: InferenceSpec; // versioned models + prior resolver affects science: yes, with uncertainty
visual: VisualSpec; // composer, LLM, texture pipeline affects science: NO
}The LLM service holds no credential and no schema field that can reach ScienceSpec.
Containment is structural, not a system prompt — and there is a test suite that proves it.
Standing on a planet 40 light-years away spans ~27 orders of magnitude. float32 gives ~7
significant digits and float64 gives ~16, so no single linear coordinate system survives
it: subtracting two huge, nearly-equal doubles annihilates the low bits.
Aether never forms the huge numbers. Every frame stores its position relative to its
parent, and two frames are related by walking to their lowest common ancestor. A one-metre
offset survives at 40 light-years to better than a nanometre — while the naive
absolute-subtraction approach loses the metre entirely. Both are asserted in
packages/coordinates/src/precision.test.ts.
git clone https://github.com/Hussaincodes01/Aether-Space-Program.git
cd Aether-Space-Program
pnpm install
pnpm dev # http://localhost:5173Requires Node 22 or newer and pnpm 9+. The floor is enforced rather than suggested:
.npmrc sets engine-strict=true, so pnpm install fails immediately and by name on an
older Node instead of producing a tree that breaks later in a confusing place.
nvm use # reads .nvmrc
node --version # must be >= 22
corepack enable # pnpm at the exact version in packageManager| Command | What it does |
|---|---|
pnpm dev |
Run the client at http://localhost:5173 |
pnpm check |
Lint + typecheck + test + build — exactly what CI runs |
pnpm lint |
ESLint across every package and the client |
pnpm format |
Rewrite with Prettier |
pnpm format:check |
Fail if anything is unformatted (CI runs this) |
pnpm typecheck |
tsc --noEmit everywhere |
pnpm test |
Vitest across every package and the client |
pnpm build |
Build every package and the web client |
pnpm build:web |
Client only |
Controls: W A S D move · R / F up and down · Shift boost · mouse wheel throttle ·
Esc release the mouse.
Panels: C catalog · P provenance · B Solar System bodies · K interstellar map ·
M missions · G precision overlay · U hide everything and fly.
Watch the local coords readout in the HUD. It stays small no matter how far the camera is from the origin — that is frame rebasing doing its job. If it ever climbs past ~1e12 m, the f32 upload will start to jitter and something has broken.
pnpm dev puts you in low Earth orbit, on today's date. Every planet sits on its real
Keplerian orbit, advanced from the JPL approximate elements — orbital periods reproduce the
published values to better than 0.1%, and to 0.01% for the terrestrial planets.
That starting point is not sentiment. The Solar System is the only place the generator can be checked against ground truth. Nobody has resolved an exoplanet surface, but we have maps of Mars. So the test suite asserts that feeding real parameters into the operator gates reproduces worlds we have actually visited:
| World | What the gates must produce | Why |
|---|---|---|
| Moon | Craters. No wind, water or ice erosion. | Airless, 0 Pa |
| Mars | Dunes and craters, but no rivers today | 6 mbar sustains the largest dune fields known; liquid water is not stable |
| Earth | Rivers, dunes and ice | 1 bar, stable water, 71% ocean |
| Jupiter | No solid surface at all | 1326 kg/m³ — nothing to stand on, and nothing to hold a crater |
If Mars came out as an ocean world, the rules would be wrong — and no amount of exoplanet work would ever have revealed it. Every exoplanet this project renders inherits its credibility from whether Mars looks like Mars.
All seven TRAPPIST-1 planets orbit closer than Mercury. NASA/JPL-Caltech, public domain.
Everything below is in the repository, tested, and building. 459 tests across twelve
packages; pnpm check runs lint, typecheck, tests and the full build.
catalog seed -> SI + provenance -> mission evidence gate -> physics derivations
-> inference for the gaps -> terrain hypothesis -> visual parameters
buildPlanetModel(systemId, planetId, evidenceProfile) produces every number the renderer
uses, and each one can name its own origin. See
docs/catalog-ingestion.md.
- ✅ f64 hierarchical frames, floating-origin camera, rebasing with hysteresis
- ✅ Typed catalog — TRAPPIST-1 (all seven), Proxima Centauri b, Sun/Earth/Mars. Every value names its unit and a source that must resolve, or ingestion fails
- ✅ Earth as the pipeline's own unit test — density 5514 kg/m³, gravity 9.80 m/s², escape velocity 11.19 km/s, insolation 1361 W/m², equilibrium temperature 254 K
- ✅ Physics registry — equilibrium temperature, insolation, density, gravity, escape velocity, scale height, Jeans parameter, Kopparapu habitable zone, blackbody colour
- ✅ Three-layer contract with structurally enforced LLM containment
- ✅ Resolver enforces its own contract — dimensional compatibility and physical domain checked on every value, rejections reported rather than swallowed
- ✅ Terrain validates by construction — one entry point that cannot be bypassed, typed errors in strict mode, unknown operators and parameters recorded
- ✅ Terrain operators with physics gates and height evaluation — craters with rims and ejecta, shield volcanoes, fracture networks, dune fields, valley incision, ice caps
- ✅ Provenance inspector (P) — six badges, unknowns rendered rather than dropped, processes ruled out listed beside those present
- ✅ Catalog panel (C) with per-planet measured/unknown counts under the current missions
- ✅ Precision overlay (G) — frame, local coordinates, rebase flag, f32 limit warning
- ✅ Rayleigh/Mie atmospheric scattering, coefficients derived from mean molecular weight
- ✅ Real Solar System — 33 bodies: the Sun, eight planets, five dwarf planets and 19 moons, ring systems. Every orbit validated against published sidereal periods
- ✅ Real sky: 109,389 stars at measured parallax distances from HYG
- ✅ Deep sky: 12,160 galaxies, clusters and nebulae from OpenNGC
- ✅ Every confirmed exoplanet: 6,309 planets in 4,722 systems, mapped
- ✅ Interstellar travel: search any of the 4,722 systems and go there
- ✅ Typed errors surfaced in the UI, plus a WebGL 2 pre-flight check
These are real and stated rather than buried:
| Limitation | Detail |
|---|---|
| Depth partitioning is opt-in | ?depth=partitioned works and is tested, but four full scene passes have not been measured on real hardware from this environment. The default remains the single-pass logarithmic buffer, which has been looked at on a screen. See docs/rendering-precision.md |
| Reversed-Z is built, not wired | Three.js does not expose the depth clear and comparison controls to drive it through its standard path |
| The typed catalog is three systems | The 4,722-system map is separate and shallower: names, distances, positions. Only catalog systems have a full interrogable model |
| One inference prior | Chen & Kipping mass-radius. Enough for Proxima b; not a full inference layer |
| No LLM composer yet | The containment boundary is built and tested. Nothing is plugged into it |
| No browser-based screenshots in CI | Scene assembly is tested headlessly (frame parenting, ring ownership, body counts) but pixels are not |
| Terrain is uncalibrated | The gates reproduce the Solar System qualitatively. Nothing yet scores generated terrain against real Mars or Moon maps |
| Document | Covers |
|---|---|
docs/provenance.md |
The data model, unit/domain checking, the resolution chain, badges |
docs/catalog-ingestion.md |
Seed data rules, the pipeline, determinism, adding a system |
docs/mission-evidence.md |
What mission selection may and may not change |
docs/terrain-safety.md |
Gates, warnings, strict mode, why terrain cannot affect physics |
docs/rendering-precision.md |
f64 frames, the f32 limit, depth strategies |
docs/PRD.md |
The original specification |
Earth's entire relief — Everest to the Mariana Trench — is about 20 km on a 6371 km radius. That is 0.3%. At true scale a terrestrial planet is smoother than a billiard ball, and every "realistic" planet you have seen rendered is vertically exaggerated, usually by one to two orders of magnitude, silently.
Aether exaggerates too, because a perfect sphere teaches nothing. The difference is that the
multiplier is a VisualSpec value, it is displayed in the HUD next to the true relief, and
setting it to 1 shows you the honest answer.
Known limitation, recorded rather than hidden: the terrain operator set does not yet model the consequences of tidal locking, so Earth and TRAPPIST-1e compile to the most similar worlds in the catalog. There is a test pinning that distance so it cannot quietly drift downward. Adding a tidal-locking operator is the top item in the backlog.
New contributors: start with CONTRIBUTING.md. It explains the three registries, how to add a feature by writing exactly one file, and the two invariants that must never break.
The short version: adding a physics formula, a terrain operator or a mission means creating one file and adding one export line. No switch statements, no dispatch tables.
Imagery is NASA/JPL-Caltech, public domain. The Hertzsprung–Russell diagram and the mission timeline are generated for this project.
Scientific method draws on the NASA Exoplanet Archive, the MALBEC radiative-transfer
intercomparison, the Kopparapu habitable-zone parameterisation, Zeng et al. mass–radius
relations, and published work on surface degeneracies in reflected-light observations. Full
citations are in docs/PRD.md.
This project uses NASA data and imagery. It is not endorsed by or affiliated with NASA, ESA, or any space agency.
Every export and screenshot carries this, burned into the framebuffer: Data-grounded artistic reconstruction. Geography is procedurally generated and is not an observation.