Transportation studies often develop separate machine learning models for individual prediction tasks, such as mode choice, trip purpose, destination prediction, or travel demand estimation. While effective for specific applications, these task-specific models fail to leverage the common patterns that exist across different aspects of human travel behaviour.

This ongoing research aims to develop a Transportation Foundation Model capable of learning a shared representation of transportation data that can support multiple prediction tasks within a single deep learning architecture. Rather than training independent models for each problem, the objective is to create a unified model that understands the underlying relationships between traveller characteristics, trip attributes, and transportation systems.

The proposed framework investigates transformer-based architectures, representation learning, and multi-task learning techniques to build reusable embeddings that generalise across a wide range of transportation applications. The long-term vision is to enable a single model to perform predictions such as travel mode choice, trip generation, destination choice, trip purpose classification, travel demand estimation, and other transportation-related tasks without requiring separate models for each objective.

By learning from diverse transportation data and sharing knowledge across tasks, the model aims to improve prediction accuracy, reduce training costs, and provide a more flexible foundation for future intelligent transportation applications.

This research is currently under development. As the project progresses, additional information regarding the model architecture, training methodology, experimental evaluation, and research findings will be published on this website.