Reconstructing urban mobility from the built environment

Editorial by Qi R. Wang: “Understanding how people move through cities is fundamental to epidemic preparedness, transportation planning, and climate policy. Yet for most of the world’s cities, particularly across the Global South, reliable mobility data simply does not exist. The traditional approach of conducting household travel surveys is expensive and slow; passively collected data from mobile phones, while transformative in data-rich countries, remains scarce where digital infrastructure is limited. This asymmetry creates a troubling paradox: the cities most in need of mobility-informed planning are precisely those with the least data to support it. For nearly eight decades, the gravity model has served as the default tool for filling this gap. The model estimates travel flows between locations as proportional to their populations and inversely proportional to the distance separating them. Its elegance lies in its simplicity: only population and distance are required. But that simplicity comes at a cost. The gravity model captures broad flow distributions while struggling to accurately estimate travel between specific pairs of neighborhoods. In recent years, deep learning approaches such as Deep Gravity have improved predictive accuracy by incorporating richer features of the built environment, but they introduce a new problem: overfitting to observed patterns and an inability to generalize to cities where no mobility observations exist. The field has been caught between interpretability and accuracy, between transferability and expressiveness (Fig. 1). In this issue of Nature Computational Science, Jinming Yang and colleagues introduce neuroGravity, a physics-informed deep learning framework that bridges this divide…(More)”.