LLM4Delay: Flight Delay Prediction via Cross-Modality Adaptation of Large Language Models and Aircraft Trajectory Representation

2025-06-09 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Aviation & Aerospace · Depth: Expert, extended

Summary

LLM4Delay is a lightweight, multimodal flight delay prediction framework designed for air traffic controllers, focusing on post-terminal delays at destination airports. It integrates aircraft trajectory data, represented by a pre-trained ATSCC encoder, with textual aeronautical information, including flight details, METAR/TAF weather reports, and NOTAMs. The framework adapts trajectory embeddings into the language modality via a cross-modality adaptation network, feeding them into a frozen large language model backbone, such as LLaMA-3.2-1B-Instruct (under 1.5 billion parameters). Evaluated on 2022 data from Incheon International Airport, the model consistently achieves sub-minute prediction error, demonstrating an average error between 0.9214 and 2.7395 minutes. This approach supports real-time, second-by-second updates, enhancing operational efficiency.

Key takeaway

For Air Traffic Controllers monitoring terminal area operations, this framework offers a significant upgrade to real-time delay prediction. You can leverage its multimodal integration of trajectory and textual data to obtain second-by-second delay updates, improving situational awareness. This approach, with its sub-minute error, provides more precise estimates than traditional methods, enabling proactive management of air traffic and reducing reliance on average delay assumptions.

Key insights

Cross-modality adaptation of trajectory data with LLMs significantly enhances real-time flight delay prediction by integrating diverse aeronautical contexts.

Principles

• LLM linguistic understanding is crucial for aviation context interpretation.
• Integrating comprehensive multimodal data significantly boosts prediction accuracy.
• Freezing pre-trained models enables efficient, lightweight training.

Method

The framework uses a pre-trained ATSCC encoder for trajectory representation, a lightweight MLP for cross-modality adaptation to LLM embedding space, and a frozen LLM backbone with a regression head.

In practice

• Integrate ADS-B trajectory data with textual flight information for enhanced ATC tools.
• Deploy lightweight LLMs (e.g., LLaMA-3.2-1B-Instruct) for real-time, sub-minute delay forecasts.
• Utilize raw METAR/TAF/NOTAM text directly, avoiding separate decoders.

Topics

• Flight Delay Prediction
• Large Language Models
• Multimodal AI
• Air Traffic Management
• Trajectory Representation
• Cross-Modality Adaptation
• Incheon International Airport

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Multilingual and Multimodal LLMs in the Wild: Building for Low-Resource Languages — Multimodal LLMs must integrate text, speech, and vision for low-resource languages, moving beyond English-centric benchmarks.
• CLEAR: Cognition and Latent Evaluation for Adaptive Routing in End-to-End Autonomous Driving — CLEAR achieves efficient, multi-modal autonomous driving by combining single-step generative planning with LLM-driven cognitive reasoning.
• Large Language Models in Transportation Systems Management and Operations: From Text Reasoning to Multi-modal Decision Support — LLMs and MM-LLMs offer a new mechanism for integrating diverse transportation data into operator-facing decision support systems.
• Lazy or Efficient? Towards Accessible Eye-Tracking Event Detection Using LLMs — An LLM-driven pipeline automates eye-tracking event detection, making complex analysis accessible without specialized programming.
• Bridging the Last Mile of Time Series Forecasting with LLM Agents — LLM agents can bridge the gap between statistical time series forecasts and business-ready predictions by integrating weakly structured contextual information.
• Learning, Fast and Slow: Towards LLMs That Adapt Continually — Combining slow parameter updates with fast in-context learning improves LLM adaptation and reduces forgetting.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.