Skip to content

Home

Neural ODE Flight Dynamics

A physics-guided Neural Ordinary Differential Equation (Neural ODE) framework for aircraft flight dynamics.

Get Started View Examples

🎯 At a Glance

node-fdm bridges the gap between deep learning and aeronautics. It allows you to compose hybrid dynamical models by stacking physical principles, analytical features, and neural networks.

The diagram below illustrates the standard architecture used for ADS-B data (OpenSky 2025), where an analytical layer pre-processes physical features before feeding them into a neural network:

graph LR
    subgraph Inputs ["System Inputs"]
        direction TB
        X((State x))
        U((Control u))
        E((Context e))
    end

    subgraph Core ["Core Blocks"]
        direction LR
        B1[Analytical Layer]
        B2[Neural Net Layer]
    end

    subgraph Solver ["Temporal Integration"]
        direction TB
        DX((Derivative dx/dt))
        ODE[ODE Solver]
    end

    %% Connexions (Ordre strict pour l'index linkStyle)
    %% Index 0, 1, 2
    X --> B1
    U --> B1
    E --> B1
    %% Index 3
    B1 --> B2
    %% Index 4
    B2 --> DX
    %% Index 5
    DX --> ODE

    %% Index 6 : Feedback Loop (Cible pour le style rouge)
    ODE -.->|Loss| X

    %% Styles
    classDef cInput fill:#9ECAE9,stroke:#333,stroke-width:2px,color:black
    classDef cControl fill:#FF9D98,stroke:#333,stroke-width:2px,color:black
    classDef cContext fill:#88D27A,stroke:#333,stroke-width:2px,color:black
    classDef cAnalytics fill:#F2CF5B,stroke:#333,stroke-width:2px,color:black
    classDef cNeural fill:#83BCB6,stroke:#333,stroke-width:2px,color:black
    classDef cDerivative fill:#D6A5C9,stroke:#333,stroke-width:2px,color:black

    class X cInput
    class U cControl
    class E cContext
    class B1,ODE cAnalytics
    class B2 cNeural
    class DX cDerivative

    %% Application du style rouge sur le lien d'index 6
    linkStyle 6 stroke:red,stroke-width:2px,stroke-dasharray: 5 5,color:red

Key Capabilities

  • Reconstruct Trajectories: Generate coherent flight paths from ADS-B or QAR data.
  • Physics-Aware: Simulate behavior using latent dynamics constrained by aeronautical laws.
  • Ready-to-Use: Includes architectures for OpenSky 2025 and QAR.
  • Benchmark Ready: Compare directly against physical models like BADA.

🚀 Workflow & Navigation

Follow the pipelines mirrored in the repository layout.

âš¡ Quick Install

You can install the core package directly via pip:

pip install node-fdm
# Or for editable research mode:
pip install -e .[all]

Legal Notice

This project is intended for research purposes only.

This project is distributed under the EUPL-1.2 license with specific EUROCONTROL amendments. It must not be used as a regulatory or operational tool under any circumstances. See AMENDMENT_TO_EUPL_license.md for details.