PSC Bridges-2 Helps Train AI to Predict and Explain Airport Collision Risks

Why It Matters

World2Rules offers a scalable, explainable solution that can give pilots and controllers critical seconds to avert runway incursions, addressing a persistent safety gap in aviation. Its neuro‑symbolic design also demonstrates a viable path for AI in other high‑risk industries where transparency and reliability are non‑negotiable.

Key Takeaways

• World2Rules improved safety rule learning by 23.6% over neural-only AI
• Neuro‑symbolic approach combines interpretability with handling noisy airport data
• Bridges‑2 processed ~10 TB of FAA surface movement data for training
• System identified unreliable evidence, reducing false collision alerts

Pulse Analysis

Runway incursions remain a stubborn safety challenge, with near‑misses like the March 2026 Air Canada‑EVA incident highlighting the need for faster, more reliable decision support. Traditional AI models either excel at pattern recognition or offer human‑readable logic, but rarely both. By marrying neural networks’ ability to digest massive, noisy datasets with symbolic methods’ transparent rule structures, the neuro‑symbolic paradigm promises the best of both worlds for safety‑critical operations.

Bridges‑2, PSC’s flagship supercomputer, proved essential for this breakthrough. Its high‑throughput architecture handled the Amelia‑42 repository—approximately 10 TB of FAA surface movement logs—allowing researchers to train World2Rules at a scale unattainable on commodity hardware. The resulting model not only generated interpretable safety rules but also filtered out unreliable evidence, delivering a 23.6% accuracy boost over pure neural approaches and a 43.2% gain versus naive symbolic techniques. Such performance gains translate directly into earlier warnings and clearer explanations for air traffic controllers.

Beyond aviation, the implications are far‑reaching. Any domain that relies on massive sensor streams—autonomous vehicles, industrial robotics, or power‑grid monitoring—could adopt a similar neuro‑symbolic framework to extract actionable, explainable policies from raw data. Future work aims to incorporate temporal dynamics, turning static snapshots into evolving predictions that better capture uncertainty. As AI continues to permeate high‑stakes environments, the World2Rules‑Bridges‑2 collaboration showcases a roadmap for building trustworthy, high‑performance systems that keep both people and assets safe.