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

Thaweerath Phisannupawong, Joshua Julian Damanik, Han-Lim Choi · 2025

Read Flight Delay Prediction via Cross-Modality Adaptation of Large Language Models and Aircraft Trajectory Representation by Thaweerath Phisannupawong, Joshua Julian Damanik, Han-Lim Choi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23636.txt.

Introduction

Thaweerath Phisannupawong and two co-authors present LLM4Delay, a framework that predicts flight delays from the viewpoint of air traffic controllers watching aircraft enter the terminal area near an airport. Delays expose inefficiency in the air traffic system, and better forecasts help controllers act early. The framework feeds a large language model both text and flight path data at once.

On the text side go flight records, weather reports, and aerodrome notices; on the other side, several aircraft trajectories that capture the state of the airspace. A cross-modality step projects each trajectory into the language model's space, so the two kinds of context blend at the level of single flights. LLM4Delay beat existing air traffic frameworks and earlier time-series-to-language methods. Because predictions refresh as new information arrives, the authors see real operational value for live traffic management.

Catalog record

CREATORThaweerath Phisannupawong, Joshua Julian Damanik, Han-Lim Choi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE95 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-23636

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Flight Delay Prediction via Cross-Modality Adaptation of Large Language Models and Aircraft Trajectory Representation Thaweerath Phisannupawong, Joshua Julian Damanik, Han-Lim Choi Abstract Flight delay prediction has become a key focus in air traffic management, as delays highlight inefficiencies that impact overall network performance. This paper presents a lightweight large language model-based multimodal flight delay prediction, formulated from the perspective of air traffic controllers monitoring aircraft delay after entering the terminal area. The approach integrates trajectory representations with textual aeronautical information, including flight information, weather reports, and aerodrome notices, by adapting trajectory data into the language modality to capture airspace conditions. The experiments show that the model consistently achieves sub-minute prediction error by effectively leveraging contextual information related to the sources of delay, fulfilling the operational standard for minute-level precision. The framework demonstrates that linguistic understanding, when combined with cross-modality adaptation of trajectory data, enhances delay prediction. Moreover, the approach shows practicality and potential scalability for real-world operations, supporting real-time updates that refine predictions upon receiving new operational information. keywords Flight Delay Prediction ,Multimodal Models ,Large Language Models ,Trajectory Representation ,Cross-Modality Adaptation † † credit: Conceptualization, Data Curation, Formal Analysis, Methodology, Software, Writing - original draft † † credit: Data Curation, Software, Writing - review and editing † † credit: Investigation, Supervision, Validation, Writing - review and editing † † corresponding: Corresponding author † † note: This work was supported by the Korea Agency for Infrastructure Technology Advancement (KAIA) Grant funded by the Ministry of Land, Infrastructure and Transport under Grant 22DATM-C163373-02. The work of Thaweerath Phisannupawong was supported by the Hyundai Motor Chung Mong-Koo Foundation Global Scholarship. 1 Introduction The volume of air traffic grows in parallel with economic development. This growth can result in traffic density that exceeds the capacity of current air traffic management (ATM) systems, compromising both safety and operational efficiency [ 1 , 2 ] . Such inefficiencies lead to flight delays that increase airline operating costs, burden airport coordination, disrupt passenger itineraries, and create network-wide ripple effects. The Federal Aviation Administration (FAA) defines delay as the excess time duration incurred during flight operations. These delays are categorized into five operational phases, namely gate delay, taxi-out delay, en-route delay, terminal delay, and taxi-in delay [ 3 ] . Delays can also be classified by cause, including carrier delays, weather delays, security delays, and airspace-related delays [ 4 ] . The delays across operational phases and causes originate from different parts of the ATM system. Understanding relevant factors is essential for accurate, phase-specific delay prediction. This study focuses on delays after the terminal phase, which are of particular interest to air traffic controllers (ATCs), as monitoring and mitigating delays within this phase is their direct responsibility. In ATM operations, the ATCs monitor trajectory data and textual aeronautical information, which belong to different modalities. Although these data are comprehensible and human-interpretable, managing, inferring, and making decisions from large volumes of such multimodal information can still be overwhelming for ATCs. The growing public availability of large-scale aviation data has enabled the development of data-driven models to assist ATCs in these tasks, enhancing autonomy while preserving effective human-machine interaction. Textual information, such as flight schedules, weather reports, and temporary operational notices, provides essential context for aviation operations and is nowadays accessible via many public databases. Aircraft trajectory data, which describes aircraft movement and operational behavior within controlled airspace, can be extracted from recorded Automatic Dependent Surveillance-Broadcast (ADS-B) databases. Integrating and understanding these contexts is important for capturing potential delay-related factors in the airspace. Since the operational context comprises information in various modalities, each requires distinct handling techniques, while their combined use necessitates effective data integration. For textual data, large language models (LLMs) have recently gained recognition as powerful models capable of understanding natural language. Pretrained on large-scale language corpora, LLMs can often interpret aviation-related text, such as flight descriptions and aeronautical information. On the other hand, trajectories are complex, high-dimensional time series that can be represented more compactly, helping downstream models perform better even without further fine-tuning. However, trajectory data is not the type of input recent LLMs are trained to handle directly; therefore, integrating such data with LLMs requires additional adaptation techniques. This paper’s proposed method incorporates two pre-trained models, an LLM and a trajectory representation model, using a simple yet effective adaptation technique to integrate their embeddings. We formulate flight delay prediction as a multimodal regression problem, where both textual and trajectory data serve as context. The contributions of this paper are as follows: • We propose a language-trajectory multimodal framework for predicting delays beyond the terminal phase, incorporating comprehensive operational context by integrating textual aeronautical information and aircraft trajectories, leveraging an LLM and a trajectory representation model for their respective modalities. • This paper uncovers that the linguistic understanding embedded in pretrained LLMs is crucial and can be effectively transferred and adapted to the flight delay prediction task, highlighting their potential for efficient multimodal inference without requiring computationally intensive full fine-tuning. • We demonstrate that incorporating a wide range of aeronautical information into the operational context leads to more accurate delay predictions, as the model can account for contextual factors that describe the airspace, thereby extending the time an aircraft spends within the area. • Our framework supports real-time delay monitoring by processing continuously updated inputs, such as surveillance broadcast signals, enabling second-by-second updates that are useful for operators who require more frequent real-time updates as the aircraft approaches the airport. 2 Related Works 2.1 Delay Prediction Flight delay prediction has long been an area of interest in the ATM community. Early studies employed tabular-based approaches, in which flight details (e.g., carrier, route, schedule, delay history) and contemporaneous weather data (e.g., visibility, wind, temperature) were extracted into structured tables and used as fixed-dimensional numerical or categorical features for machine learning models. Delay prediction can be formulated as a classification problem, where the goal is to determine whether a delay will occur. For example, prior studies have employed logistic regression [ 5 ] and a modified Random Forest model [ 6 ] . Meanwhile, several studies [ 7 , 8 , 9 , 10 ] have compared various non-neural machine learning models for delay classification. Integrating traffic count features, a model presented in [ 11 ] predicts departure delays using linear regression, support vector machines (SVM), Multilayer Perceptron (MLP), and LightGBM. The delay prediction can also be formulated as a regression problem that predicts delay duration in minutes. For example, SVM [ 12 ] and an MLP using only flight information [ 13 ] have been applied to predict delay duration. The work in [ 14 ] compares several models for predicting delay distributions based on flight features. Subsequently, studies [ 15 , 16 , 9 ] compare various models for delay regression. Many works have incorporated historical data for delay classification in sequence. For instance, the methods proposed in [ 17 , 18 ] both employ long short-term memory (LSTM) models, utilizing flight information and weather data to classify next-day delays. However, the day-level granularity in both limits their real-time applicability. Complex network models, such as the LSTM-integrated random forest model [ 19 ] and the Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) [ 20 ] , were

Read the complete text