AERO-FOUNDATIONAEROSPACE SUPERINTELLIGENCE · RESEARCH PROGRAM
aero-foundation
Domain model · Tool-verified rewards

An aerospace model that learns from the tools engineers trust.

aero-foundation specializes an open-weights language model for flight control, aerodynamics, and handling qualities. It is trained with continued pretraining, then reinforcement learning in which every reward comes from an engineering tool that computes a checkable answer.

VIEW ON GITHUB HOW IT WORKS
Phase 0 · founding Apache-2.0 Ground tool · not airborne DO-330 qualification path
Mission

Engineering judgment, shaped by engineering checks.

Aerospace already has rigorous ways to tell a right answer from a wrong one. General-purpose models rarely use them, so they produce designs that read well and fail the first check. We train on the checks themselves.

01

Flight control design

Gains and architectures that meet stability margins, bandwidth, and time-response requirements.

02

Aerodynamic databases

Set up and run analyses, then assemble coefficient tables and derivatives that agree with the tools.

03

Handling qualities

Compute modal characteristics and grade them against MIL-F-8785C and MIL-STD-1797A flying-qualities criteria.

04

Requirements traceability

Link requirements to design artifacts and verification cases, in the style ARP4754B and DO-178C expect.

Approach

Propose. Check. Reward. Update.

Continued pretraining on a license-clean aerospace corpus gives the model the vocabulary. Reinforcement learning gives it judgment: the model proposes, a sandboxed tool checks, and the check is the reward. No learned reward model sits between the physics and the gradient.

Training loop: an aerospace corpus and an open-weights base model feed continued pretraining, producing the policy. The policy proposes a design, engineering tools check it, the result becomes a reward, and an RL update improves the policy. Tool inputs and outputs are written to an evidence log. Aerospace corpuslicense-tagged Base modelopen weights Continuedpretraining Policy Engineering toolsJSBSim · AVL · XFOIL Reward evidence log propose check RL update (e.g. GRPO) Training loop: corpus and base model feed continued pretraining, producing the policy; the policy proposes, tools check, the check becomes a reward, and an RL update improves the policy. Corpuslicense-tagged Base modelopen weights Continued pretraining Policy Engineering toolsJSBSim · AVL · XFOIL Reward evidence log propose check RL update
Fig. 1 · Training loop. Tool versions are pinned and hashed so any reward can be reproduced.
Verifiers & tools

Rewards computed, not judged.

Each verifier runs in a sandbox with timeouts and input validation. A tool that fails to converge returns zero reward, never a crash or a default pass. Where two tools overlap, we cross-check them to expose blind spots.

ToolChecksExample reward
JSBSimNonlinear 6-DOF flight: trim, time responses, closed-loop behaviorStep response meets rise-time and overshoot specs
AVLVortex-lattice aero: lift, induced drag, moments, stability derivatives, neutral pointGeometry trims at the target CL with static margin in band
XFOIL2-D airfoil polars, transition, pressure distributionAirfoil meets a cl/cd target at the given Reynolds number
OpenVSPParametric geometry, VSPAERO analysisValid geometry matching area and span constraints
python-controlPoles, gain and phase margins, bandwidth, step metricsLoop meets margins with the specified crossover
HQ criteriaMIL-STD-1797A / MIL-F-8785C modal criteria: short period, CAP, Dutch roll, roll mode, spiral, bandwidthLevel 1 in every required flight phase
Certification philosophy

The model proposes. Qualified tools decide.

aero-foundation is a ground-based engineering tool, not airborne software. The long-term path is tool qualification under DO-330, within programs that follow ARP4754B, DO-178C, and DO-331.

The model is a proposer

Non-deterministic and hard to specify, so its outputs are treated as unverified until checked. We do not plan to claim certification credit for the model itself.

The checkers are qualifiable

Deterministic, version-pinned tools whose behavior can be specified, tested, and placed under configuration management. Credit, if ever claimed, comes from independent verification of outputs.

  • ARP4754B
  • DO-178C
  • DO-330
  • DO-331
  • MIL-STD-1797A

Traceability is built in from day one: data provenance manifests, training configurations, weight hashes, verifier hashes, and per-task evidence logs. Qualification is granted within a specific certification project by the relevant authority; nothing here is qualified today. This is engineering research, not regulatory advice.

Roadmap

Evidence before scale.

Each phase ends with a checkable exit criterion. Compute estimates in the experiment plan are stated as assumptions.

Phase 0 · now

Foundations

Verifier harness, task bank, eval suite, provenance pipeline, base-model baselines.

ExitReproducible baselines for two or more open models from one command.
Phase 1

RL with tool rewards

RLVR on a small base model with Phase 0 verifiers; no continued pretraining.

ExitSignificant gain on held-out tool-graded tasks, no general regression.
Phase 2

Continued pretraining

Aerospace corpus with license tagging, dedup, and decontamination.

ExitDomain gains without catastrophic forgetting.
Phase 3

CPT + RL at scale

Combined stages; closed-loop JSBSim, aero databases, traceability verifiers.

ExitBeats the strongest prompted baseline on the full suite.
Phase 4

Qualification hardening

TOR and TQP drafts for the verifier tools; automated evidence.

ExitEvidence-package dry run against DO-330 objectives.
Research

Start with what exists.

Every source in these documents was opened and checked. Nothing is invented; everything is paraphrased.

Status

Founding phase. No model yet.

The repository holds research and planning documents. There is no trained model and no capability claim. Work is tracked in GitHub issues under the Phase 0 milestone, on four tracks: data, verifiers, eval, and certification.

OPEN ISSUES CONTRIBUTE
phase0 · foundations
modelnone trained
claimsnone
licenseApache-2.0
scopeground tool
Citation

Cite this work

@software{aero_foundation_2026,
  title   = {aero-foundation: an aerospace engineering model
             trained with tool-verified rewards},
  author  = {{Aerospace Superintelligence contributors}},
  year    = {2026},
  url     = {https://github.com/aerospace-superintelligence/aero-foundation},
  license = {Apache-2.0}
}