Beyond Euclidean Zoning: Machine Learning Reveals Empirical Patterns of Land Use Compatibility

Despite over a century of zoning practice, land use compatibility — the degree to which adjacent land uses coexist without generating negative externalities — remains assessed through expert judgment and inherited assumptions rather than empirical evidence of urban outcomes. Given that compatibility standards underpin billions of dollars in annual zoning decisions, their empirical validity carries direct consequences for urban livability. This study presents a machine learning framework that shifts compatibility assessment from subjective prescription to data-driven discovery. Applying gradient boosting regression to 134,863 census blocks across New York, Los Angeles, Chicago, and Houston, the model predicts composite compatibility scores derived from six dimensions — functional diversity, activity complementarity, conflict minimization, accessibility synergy, socioeconomic integration, and infrastructure efficiency — using 32 features capturing land use composition, built environment morphology, transportation infrastructure, and socioeconomic context. The model achieves strong predictive performance (R² = 0.895, RMSE = 0.087), establishing that compatibility follows systematic, learnable patterns. Land use entropy dominates prediction (63.7% feature importance), providing robust evidence that functional mixing enhances neighborhood quality — directly challenging Euclidean use-separation zoning. Context-specific compatibility matrices reveal that 45 of 55 land use pairings are empirically estimable in transit-oriented contexts versus only one in auto-dependent suburbs, demonstrating that compatibility is locally conditioned rather than universally fixed. Most pairings fall within a moderate range, with mixed-use designations consistently scoring highest. The study identifies empirically derived attribute thresholds for high-compatibility neighborhoods, offering planners measurable benchmarks — including optimal entropy ranges, building height and coverage parameters, and land use composition targets — to replace arbitrary development standards.