Travel mode choice plays a fundamental role in transportation planning, directly influencing travel demand forecasting, infrastructure investment, and policy evaluation. Traditional mode choice models often rely on statistical assumptions that may struggle to capture the complex relationships between traveller characteristics, trip attributes, and transportation networks.

This ongoing research investigates the application of modern deep learning techniques to travel mode choice prediction using large-scale travel survey data collected in Tehran. The project explores transformer-based architectures capable of learning complex feature interactions while providing a flexible framework for modelling travel behaviour.

In addition to transformer models, the research examines the integration of Physics-Informed Neural Networks (PINNs) to incorporate transportation domain knowledge into the learning process. By embedding theoretical or physical constraints within the model, the objective is to improve prediction accuracy while enhancing model robustness and generalisability.

The project also studies advanced representation learning techniques for heterogeneous transportation data, with the goal of developing models that can better understand traveller behaviour and support future transportation planning applications.

This research is currently in progress. Additional technical details, experimental results, and publications will be added as the project reaches key milestones.