An autonomous launch-vehicle design engine whose physics is validated against rockets that actually flew.
Give AEROS a mission — "5 tonnes to 500 km" — and it designs a complete multistage launch vehicle: engine selection, staging, propellant loads, tank structure, and dimensions. Then it proves the design works by flying it in a 3-DOF trajectory simulator, and it shows you every engineering decision it made, with the reasoning.
No neural network guesses a rocket here. Every number comes from physics you can check.
$ aeros design "5000 kg to 500 km"
Mission: 5.00 t to 500 km (launch lat 28.5 deg)
AEROS design: GLOW 78.1 t, lift-off T/W 2.43, ideal dv 9018 m/s
Stage 1: 1x RS-25, prop 41.6 t, dry 8.98 t, burn 81 s
Stage 2: 2x RL10C-1, prop 20.0 t, dry 2.02 t, burn 431 s
VERIFICATION: orbit achieved (perigee 485 km, max-q 47 kPa, max accel 6.5 g)
DECISIONS:
[delta-v budget] 8958 m/s
why: circular velocity at 500 km plus standard gravity/drag/steering
losses, minus Earth-rotation credit at 28.5 deg latitude
[architecture] 2-stage, 1x RS-25 + 2x RL10C-1, 2.86 m core
why: minimum-GLOW architecture out of 30 sized candidates; closest
competitor: Raptor 2/RL10C-1 at 86.3 t GLOW
...
And now the same design can be exported as concept CAD:
$ aeros cad "5000 kg to 500 km" --out cad_out/
CAD export:
obj: cad_out/vehicle.obj
stl: cad_out/vehicle.stl
scad: cad_out/vehicle.scad
manifest: cad_out/vehicle_manifest.json
review_png: cad_out/vehicle_cad_review.png
review_json: cad_out/vehicle_cad_review.json
Anyone can generate a rocket-shaped drawing. The question is whether the physics engine that sized it can be trusted. AEROS answers that by flying real vehicles, built from their published stage masses and engine data, through the exact same simulator, and comparing predicted payload capacity with the operators' published figures:
| Vehicle | Published payload | AEROS prediction | Error |
|---|---|---|---|
| Falcon 9 Block 5 (expendable, 200 km LEO) | 22.8 t | 24.95 t | +9.4% |
| Rocket Lab Electron (200 km LEO) | 300 kg | 333 kg | +11% |
| Saturn V two-stage (Skylab, 435 km / 50°) | 77.1 t flown | mission reproduced, capacity ≈ 121 t | ✓ |
The simulated flights also reproduce flight-measured quantities that were never fitted: maximum dynamic pressure comes out at 29–36 kPa across all three vehicles (Falcon 9 flies ~33 kPa in reality), and staging velocities and accelerations land in the flown ranges.
About the +10% bias: it is consistent and explainable. The simulator flies ideal guidance with no flight-performance reserve, no wind dispersions, and no operator margin — so it should slightly exceed published (conservative) figures, and it does, by a similar amount on a 13 t rocket and a 550 t one. We consider honest, characterized bias more valuable than a suspiciously perfect fit.
Reproduce it yourself: aeros validate (takes a while — it re-optimizes ascent steering at every bisection step).
aeros/atmosphere.py— US Standard Atmosphere 1976, seven-layer implementation, unit-tested against the published tables.aeros/engines.py— database of real, flown engines (Merlin, Raptor, RS-25, F-1, J-2, RL10, Rutherford) with exact nozzle back-pressure relationF(p) = F_vac − A_e·p; the model's sea-level Isp must reproduce each engine's published value within 3% (unit-tested).aeros/trajectory.py— 3-DOF ascent dynamics with adaptive RK45: air-relative aerodynamics over a rotating Earth, Mach-dependent drag, q·α structural steering envelope, automatic max-q throttle bucket, two-burn orbital insertion, differential-evolution steering optimization, payload capacity by bisection.aeros/structures.py— stage dry-mass buildup: hoop-stress-sized tank shells, engine and thrust-structure masses, calibrated to reproduce four flown stage masses within ±10% (unit-tested).aeros/design.py— the autonomous designer: enumerates engine architectures, sizes each to mass closure, optimizes staging and diameter for minimum GLOW, verifies the winner by simulation, and logs every decision with its rationale.aeros/cad.py— concept CAD generation: OBJ/STL/OpenSCAD export, explicit thrust/stage/payload interface markers, corrected clustered engine layouts, a JSON engineering manifest, and rendered CAD review sheets with automated geometry checks.aeros/validate.py— the reference vehicles and the validation harness.aeros/plots.py— vehicle drawings, ascent profiles, and one-page design verification sheets (matplotlib only).docs/space_problem_intelligence_2026.md— official NASA/ESA problem-intelligence note explaining why CAD + structured design data is the next AEROS layer.
git clone https://github.com/JISSAL-GIGI/aeros
cd aeros
pip install -e .[dev]
pytest tests/ -q
python examples/design_demo.py
aeros cad "1500 kg to 400 km" --out my_cad/
aeros design "1500 kg to 400 km" --out my_design/Requires Python ≥ 3.10, numpy, scipy, matplotlib. No GPU, no CAD kernel, no external simulation software. The first CAD layer writes OBJ/STL/OpenSCAD directly from Python and emits a review PNG plus JSON checks for stack continuity, engine clearance, interface markers, and fairing placement.
Three capabilities from the AEROS gap analysis are now built in:
Uncertainty-aware design (aeros/uncertainty.py) — Monte Carlo propagation of concept-level input uncertainty (stage dry mass ±7%, Isp ±0.7%, drag ±10%, residuals 0.3–1%) through the flight-validated simulator. Instead of one payload number you get P(orbit) vs payload, and chance_constrained_design resizes vehicles until they reach orbit at a stated confidence. For Falcon 9, the 95%-confidence capacity (20.1 t) and the deterministic capacity (24.95 t) bracket the operator's published rating (22.8 t) — quantitative evidence that published ratings embed uncertainty margin.
Cost as an objective (aeros/cost.py) — TRANSCOST-class first-unit CERs (stage dry mass and engine power laws, engine-cluster learning curve, commercial-practice factor). The GLOW-vs-cost trade across all sized architectures shows the mass-optimal vehicle is not the cost-optimal one:
Fast mode (aeros/fastopt.py) — warm-started steering refinement with a short-global-solve fallback. Measured on Electron: capacity 328 kg vs 333 kg for the full differential-evolution search (−1.5%) at ~4× less wall time; successful warm steps cost ~2 s vs 25–50 s for a global solve.
We are explicit about scope, because trust requires it:
- 3-DOF, planar, spherical Earth. No 6-DOF attitude dynamics, no winds, no dispersions, no J2. Adequate for conceptual capacity studies (as the validation shows), not for mission design.
- Conceptual-level structures. Mass-estimating relationships calibrated on flown stages — not FEA. The ±10–15% structural accuracy bounds the design fidelity.
- Concept CAD, not production CAD. The CAD exporter creates inspectable stage/fairing/engine/interface geometry, a structured manifest, and visual review artifacts; it is not yet a watertight, tolerance-controlled, fabrication-ready STEP assembly.
- Expendable two-stage designs. Reusability penalties, boost-back budgets, parallel staging (boosters), and 3+ stage optimization are on the roadmap.
- No costing. GLOW is the current objective; cost models are deliberately excluded until they can be sourced credibly.
- Published-data uncertainty. Reference stage masses carry real uncertainty (SpaceX does not publish exact dry masses); the validation errors should be read with that in mind.
AEROS is the first published layer of a larger effort — an autonomous engineering system that goes from mission statement to validated design with full decision traceability. Next layers, in order:
- Multi-objective design — Pareto fronts over GLOW / cost / reliability instead of single-objective GLOW.
- Higher-fidelity verification loop — 6-DOF trajectory, load cases, and automated FEA hand-off for the structures the conceptual sizer proposes.
- CAD interface intelligence — explicit attachment points, load paths, standard interfaces, manufacturing constraints, and optional STEP export through a CAD kernel.
- Parallel staging and reusability — boosters, propulsive recovery budgets.
- Knowledge layer — requirements parsing and design-rule extraction from the open engineering literature, feeding the same verified physics core.
MIT. Built to be checked, challenged, and extended — if you find a physics error, please open an issue; that is the entire point of publishing the validation.