Poster #8 - Analyzing Urban Mobility Using a Statistical Mechanics Approach

• In Event: Friday Poster Reception

Abstract

Background and Research Question
We develop a framework that represents traffic flows in a manner that provides practical, actionable insights, not only about roads and specific locations but more importantly, for the entire transportation network of an area. Using Seattle as a case study, we apply Physics-Inspired Machine Learning (ML) techniques that consider the interdependencies among the volume of generated, the origin and destinations of these trips, the associated mode choices and route assignments. We present findings on novel travel pattern metrics that allow for objective comparisons over time and across different geographic areas. A key step in the analytical process involves the Hodge decomposition, which enables an effective segmentation of geographic zones, the mapping of complex travel patterns, and the provision of practical insights for policy adjustments in response to changing urban travel dynamics. Our approach is sufficiently robust and comprehensive enough to explore the implications of socio-demographic changes, infrastructural developments, and external shocks such as the recent Baltimore Harbor Bridge collapse on travel behavior.

Methods
Our approach integrates data from multiple sources—LODES, GTFS, and UTD19—to construct directed mobility graphs where nodes represent geographic units and edges represent commuter flows. Spectral clustering identifies latent travel communities, while Hodge decomposition dissects flows into gradient (commuter pressure), harmonic (persistent cycles), and curl (localized loops) components in a manner that preserves the interactions across these elements. Gradient-boosted trees and SHAP analysis link travel volumes to socio-demographic predictors, while a Naive Bayes classifier estimates discrete mode choice probabilities. We validate model accuracy and policy applicability through comparisons to American Community Survey (ACS) baselines and statistical metrics such as modularity and assortativity.

Results
Our analysis uncovers distinct morning-evening ``phase transitions'' in node potentials and identifies travel communities not aligned with administrative boundaries. Gradient flows explain 87% of commuting variation, with persistent harmonic cycles (e.g., I-5 corridor) comprising 18% of total traffic. SHAP values highlight income, density, and transit access as key predictors of travel demand. The block-group-level model achieves 92% prediction accuracy---surpassing traditional tract-level baselines. High global potential variance points to systemic socio-spatial imbalances in job-housing distribution.

Conclustions
This study offers a novel, physics-informed framework for understanding cities as non-equilibrium systems. Our method provides granular and interpretable insights into traffic dynamics, enabling policymakers to target infrastructure investments, design dynamic congestion pricing, and evaluate equity in transit accessibility. By bridging urban planning, data science, and statistical mechanics, this work introduces scalable tools for evidence-based decision-making in transportation systems.

Authors

• Yangyang Wang, University of Michigan Ann Arbor
  Presenting Author

• TAYO FABUSUYI, University of Michigan Ann Arbor
  Non-Presenting Co-Author