AI Predicts Flight Delays in Under a Minute

Air travel delays cost airlines billions annually and frustrate millions of passengers, but a new AI system can predict these delays with surprising accuracy in less than a minute. Researchers from the Korea Advanced Institute of Science and Technology have developed a method that combines flight schedules, weather reports, and real-time aircraft positions to forecast delays as flights approach airports. This approach specifically helps air traffic controllers monitor aircraft entering terminal airspace, where congestion often causes cascading delays.

The key finding is that this system achieves sub-minute prediction times while maintaining high accuracy. By integrating multiple data sources—including flight plans, meteorological reports, aerodrome notices, and aircraft trajectories—the model predicts what researchers call the 'post-terminal duration.' This is the time between when an aircraft enters terminal airspace and when it actually lands, which directly contributes to total arrival delays. Experimental results show the model maintains prediction errors below one minute across various test scenarios.

The methodology uses a lightweight adaptation technique that combines large language models with specialized trajectory encoders. Rather than processing complex flight data directly, the system converts aircraft movement patterns into language-like representations that AI models can understand. This cross-modality approach allows the AI to interpret both textual information (like weather forecasts) and numerical trajectory data simultaneously. The researchers tested multiple AI backbone models and found that instruction-tuned versions performed best, with Meta's LLaMA-3.2-1B-Instruct achieving the highest accuracy across evaluation metrics.

Results from testing on 2022 flight data at Incheon Airport demonstrate the system's practical value. The model achieved mean absolute errors as low as 0.9 minutes for predicting post-terminal durations, with R-squared values exceeding 0.999 for total delay prediction—indicating near-perfect explanation of variance in the data. When researchers removed different data components in ablation tests, they found that trajectory information contributed most significantly to accuracy, followed by weather data and flight information. The system's ability to update predictions second-by-second using real-time surveillance data makes it particularly valuable for dynamic air traffic control operations.

This technology matters because it addresses a critical bottleneck in aviation efficiency. Current delay prediction methods often rely on historical patterns or limited real-time data, but this system incorporates the full context available to air traffic controllers. For regular travelers, more accurate delay predictions could mean better connection planning, reduced airport congestion, and potentially lower ticket prices as airlines optimize operations. The system's lightweight design means it could be deployed without requiring expensive hardware upgrades at control facilities.

The research acknowledges limitations, including that the model was trained and tested primarily on data from a single airport. The paper notes that performance might vary at airports with different traffic patterns or weather conditions. Additionally, while the system handles multiple data modalities effectively, its accuracy depends on the quality and timeliness of incoming data streams. The researchers suggest future work could extend the approach to other flight phases and explore integration with conflict detection systems.