Generative Digital Twin for Aviation Safety AI

Generative AI–powered digital twin that predicts and prevents unstable aircraft approaches in real time using public air traffic data.

Comment

Early Detection of Unstable Aircraft Approaches

Using Energy-State Deviations from ADS-B Trajectories

Project Story

About the Project

Unstable approaches are a major precursor to runway excursions, hard landings, and last-minute go-arounds. Despite strict stabilization criteria in commercial aviation, real-time large-scale detection remains limited, especially without proprietary flight data.

This project introduces an explainable machine learning framework that detects unstable approaches using energy-state deviations derived entirely from public ADS-B data.

The core idea:
Instead of relying on black-box flight recorder data, can we detect instability using physics-grounded features extracted from publicly accessible trajectories?

The answer turned out to be yes.

Inspiration

Commercial aviation safety relies heavily on post-flight analysis and proprietary data streams. However, global ADS-B data is openly available through networks like OpenSky.

The opportunity was clear:

  • Use physics instead of heuristics
  • Model aircraft energy states during descent
  • Detect instability 2–10 NM from runway threshold
  • Provide 2–3 minutes of early warning

The goal was to build an operationally viable early-warning system without requiring airline-owned flight data.

The Core Idea: Energy-State Deviations

An aircraft on a stabilized approach follows predictable energy dynamics:

E = 1/2 m V^2 + m g h

Where:

  • V = velocity
  • h = altitude
  • m = mass
  • g = gravitational acceleration

Instead of absolute energy (mass is unknown in ADS-B), we model relative energy-state deviations derived from:

  • Vertical descent rate
  • Speed deviations
  • Glide-path angle
  • Distance to runway threshold

Unstable approaches manifest as anomalous deviations in these energy components.

Data Engineering

We integrated multi-source public datasets:

  • OpenSky ADS-B trajectories
  • Airport runway geometries
  • METAR weather observations

From these, physics-grounded features were engineered:

  • Excess kinetic energy proxies
  • Vertical path deviation
  • Sink rate instability
  • Speed variance near threshold
  • Wind-influenced trajectory shifts

This ensured the model remained interpretable and aligned with aviation safety logic.

Model Development

We trained multiple models for comparison:

  • Logistic Regression
  • Random Forest
  • XGBoost

The final model used XGBoost due to its performance and robustness.

Performance Results

  • XGBoost AUROC: 0.9513
  • Random Forest AUROC: 0.9439
  • Logistic Regression AUROC: 0.944

Through rigorous ablation studies, performance remained strong:

AUROC = 0.9982

This demonstrated that instability detection is strongly driven by core descent dynamics rather than overly complex feature engineering.

Early Warning Capability

The system detects unstable approaches at:

  • 2–10 nautical miles from threshold
  • Providing approximately 2–3 minutes of intervention time

This time window is operationally meaningful for corrective pilot action.

Explainability & Trust

Aviation systems demand interpretability.

We used:

  • SHAP values
  • Partial dependence plots

Findings showed:

  • Vertical descent dynamics dominate risk contribution
  • Speed deviations significantly increase instability probability

This aligns directly with ICAO stabilization criteria and industry best practices.

The model does not behave like a black box, it reflects real aviation physics.

Challenges

Rare Event Detection
Unstable approaches are relatively infrequent → class imbalance issues.

Mass Not Available in ADS-B
Absolute energy cannot be computed → required normalized energy proxies.

Data Noise
ADS-B trajectories contain sampling gaps and sensor inaccuracies.

Operational Validity
Ensuring predictions align with real-world flight dynamics standards.

What I Learned

  • Physics-informed ML dramatically improves interpretability.
  • Public aviation data is more powerful than commonly assumed.
  • Explainability is non-negotiable in safety-critical systems.
  • Simplicity in feature design often outperforms overly complex pipelines.

Impact

This project demonstrates that:

  • Public ADS-B data can support safety-grade ML systems
  • Early detection of unstable approaches is feasible without proprietary data
  • Explainable AI can align with aviation regulatory standards
  • Scalable, airport-wide monitoring systems are technically achievable

It addresses a critical safety gap while remaining transparent, physics-grounded, and operationally realistic.

Vision Forward

  • Real-time airport-wide deployment
  • Integration with ATC advisory systems
  • Multi-airport global instability monitoring
  • Expansion to runway excursion probability modeling

Aviation safety does not need to wait for incidents to happen.
With energy-state modeling and explainable ML, instability can be detected, and mitigated before touchdown.

Built With

Share this project:

Updates