Cities worldwide confront a fundamental tension: how to simultaneously achieve environmental protection, social equity, and economic vitality when these objectives often conflict. Nearly 55% of the global population resides in urban areas, a proportion projected to reach 68% by 2050, intensifying pressures on municipal governments to balance competing sustainability imperatives¹. Spatial planning decisions represent the most powerful intervention point, fundamentally structuring how cities distribute environmental resources, social infrastructure, and economic opportunity². Yet configurations optimizing environmental quality typically exclude lower-income residents, while affordable neighborhoods systematically lack environmental amenities and economic opportunities. With a pattern replicated across diverse geographic and economic contexts, Chicago exemplifies this sustainability trilemma. The city's lakefront neighborhoods enjoy air quality 23% better than the city average, yet median home values exceed $450,000—pricing out 68% of Chicago households³. Conversely, the South and West sides, where 72% of residents identify as Black or Latino, experience flood risk 3.2 times higher and green space access 47% below the municipal median, despite offering more affordable housing⁴. This reveals how spatial planning decisions perpetuate systematic inequalities even within a single metropolitan region.

In addition to the exclusion of the marginalized population, traditional planning approaches struggle with the complexity of spatial distribution of resources. Comprehensive plans establish qualitative goals—"promote sustainable development," "enhance livability"—without quantifying trade-offs or identifying optimal configurations⁵. Zoning codes evolved to separate incompatible uses rather than optimize sustainability performance⁶, while master plans operate on 20-year horizons poorly suited to rapidly evolving climate targets⁷. Cost-benefit analysis typically monetizes outcomes within a single dimension, obscuring fundamental incommensurability between environmental preservation, social justice, and economic efficiency⁸. These computational limitations have historically masked systemic patterns—a cognitive barrier we have mistaken for physical impossibility. Traditional planning, lacking quantitative frameworks to navigate these trade-offs, defaults to incremental adjustments that perpetuate inequalities.

Early urban planning optimization relied on linear programming for well-defined problems like facility location but proved inadequate for multi-dimensional sustainability challenges with nonlinear relationships and competing objectives^9,10. Evolutionary algorithms offered a breakthrough by enabling multi-objective optimization. Models such as NSGA-II became the dominant approach for discovering Pareto-optimal solutions—configurations were improving any one objective necessarily degrades another¹¹. NSGA-II's population-based search maintains diverse non-dominated solutions representing optimal trade-offs¹². However, conventional methods require thousands of evaluations¹³, making it computationally infeasible to analyze all neighborhoods simultaneously, preventing city-wide pattern discovery

Bayesian optimization (BO) enables city-wide analysis through intelligent sequential sampling guided by probabilistic surrogate models¹⁴. BO constructs Gaussian Process models that predict objective values and quantify uncertainty, enabling strategic candidate selection with dramatically fewer evaluations. The qEHVI acquisition function extends BO to multi-objective problems by measuring expected improvement to the Pareto front's dominated hypervolume¹⁵. Recent algorithmic advances enable simultaneous evaluation of entire urban systems rather than pre-selected neighborhoods¹⁶. Despite these computational advantages, existing frameworks employ generic features—raw census variables and land use categories—that fail to capture domain-specific sustainability mechanisms¹⁷. Environmental models use basic metrics like "percentage tree cover" without accounting for density interactions. Social assessments rely on raw income without constructing housing cost burden indices. This feature engineering gap limits even sophisticated optimization algorithms.

Three critical gaps constrain existing frameworks: computational tractability, domain awareness, and actionable insights. Here, we integrate domain-specific feature engineering, objective-tailored predictive models, and sample-efficient multi-objective Bayesian optimization to address these limitations. We develop 10 additional composite indices grounded in urban planning theory with separate LightGBM models for each objective. Multi-objective Bayesian optimization with qEHVI discovers Pareto-optimal solutions across all 801 census tracts simultaneously. Applied to Chicago, this framework identifies concrete land use configurations and reveals preconditions enabling multi-objective success, directly supporting UN Sustainable Development Goal 11¹⁸.

This research advances urban sustainability science through four contributions. Methodologically, city-wide simultaneous analysis transforms optimization from pre-selected neighborhood evaluation to discovering systemic patterns. Empirically, novel features yield R² >0.96 across objectives, operationalizing urban planning theory within machine learning. Substantively, 59 Pareto-optimal solutions reveal fundamental trade-offs: no solution achieves > 0.50 in both environmental and economic objectives, while balanced solutions converge on 25.3% open space at 16,000–24,000 persons/km². Practically, findings identify transit accessibility and housing affordability as leverage points, with green infrastructure enabling multi-objective success, directly informing zoning reforms and capital budgets. Addressing this requires novel computational approaches capable of simultaneously evaluating environmental, social, and economic outcomes across entire urban systems.

Traditional urban planning approaches evaluate sustainability trade-offs sequentially across tens of neighborhoods—a computational constraint long mistaken for physical impossibility. Our AI-driven multi-objective optimization framework transcended this limitation by simultaneously evaluating Chicago’s 801 census tracts across three competing dimensions, discovering 59 Pareto-optimal configurations that were physically achievable but computationally invisible to conventional analysis. Most critically, these solutions cluster exclusively along century-old infrastructure corridors established by Burnham's 1909 Plan, with zero in historically disinvested South and West Chicago. This geographic pattern reveals how historic investments create path dependencies that contemporary policy struggles to overcome—constraints that remain hidden without simultaneous computational evaluation across all neighborhoods.

The AI framework successfully operationalized urban planning theory within predictive models, achieving robust predictive capability across all three sustainability dimensions (Table 1). The environmental model achieved the highest accuracy (R² = 0.985), followed by economic (R² = 0.982) and social models (R² = 0.968). All models-maintained prediction errors below 2% of the normalized scale (RMSE ≤ 0.020). These high accuracy levels validate that the framework can reliably predict sustainability outcomes for configurations never observed in historical data, enabling genuine optimization rather than mere interpolation. Evaluating all 801 tracts reveals patterns invisible to sequential analysis: zero optimal solutions exist in historically disinvested South and West Chicago.

The machine learning feature importance analysis revealed non-obvious drivers that challenge conventional planning assumptions (Table 2). Environmental outcomes were dominated by ecological infrastructure, with plant diversity contributing 45.3% of predictive power—nearly five times more influential than any other environmental factor. This dominance suggests that biodiversity serves as a master indicator of ecological health in urban contexts, synthesizing multiple ecological processes including air quality, carbon sequestration, and ecosystem resilience. Social outcomes prioritized economic opportunity, with the employment index accounting for 37.3% of predictive importance, while traditional demographic factors like education (9.7%) played secondary roles. This finding challenges the conventional planning emphasis on educational attainment as a primary social intervention point, suggesting that job availability may be a more direct lever for social outcomes. Economic predictions relied heavily on housing market dynamics (39.5%), with our engineered economic accessibility feature ranking second (9.8%), validating our domain-specific feature engineering approach. The relatively low importance of direct employment rate (4.6%) compared to housing values indicates that property markets capture cumulative economic vitality better than labor statistics alone. These patterns provide empirical validation for redirecting planning interventions toward leverage points invisible in traditional single-variable analyses.

We discovered 59 Pareto-optimal solutions spanning environmental scores 0.477–0.704, social scores 0.687–0.769, and economic scores 0.370–0.623 (Fig. 2). These represent the theoretical performance ceiling where improving one objective necessarily degrades another. The environmental-economic frontier shows the strongest negative correlation: no solution achieves > 0.50 in both simultaneously. Environmental-social objectives show strong conflict—solutions achieving environmental scores above 0.65 consistently yield social scores below 0.73, while high social performers (> 0.76) restrict environmental scores to below 0.56. Conversely, the social-economic frontier demonstrates weak positive correlation, with multiple solutions exceeding 0.75 social and 0.50 economic scores concurrently.

The three independent trials validated the optimization's robustness (Fig. 3). Trial 2 identified 28 solutions (47% of total), Trial 0 found 22 (37%), and Trial 1 contributed 9 (15%). All extreme configurations appeared within the first 30 iterations, while intermediate solutions emerged throughout the 50-iteration process, demonstrating the algorithm's efficiency in discovering the complete Pareto front structure with 110 evaluations per trial. The three trials' convergence on the same geographic locations—despite discovering solutions with distinct performance profiles—reinforces that only certain areas of the city can achieve multi-objective optimization under current conditions.

The narrow social score range (0.082) compared to environmental (0.227) and economic (0.253) range reveals social objectives are most robust to land use variations—planners have greater flexibility achieving social equity across different spatial configurations than balancing environmental and economic goals. This finding has important policy implications: social outcomes appear relatively achievable across diverse development patterns, while environmental and economic goals require precise calibration of land use configurations.

We identified four distinct archetypes—environmental leaders, social leaders, economic leaders, and balanced solutions with density spanning nearly seven-fold (4,881 − 33,445 persons/km²) (Fig. 1). Environmental leaders achieve 0.704 at just 4,881 persons/km² but sacrifice economy (0.392). Social leaders peak at 0.769 across flexible densities (15,301 − 31,796 persons/km²), suggesting well-being depends more on service provision than density. Economic leaders reach only 0.620 at 24,049 persons/km² while sacrificing environmental quality (0.490).

Critically, balanced solutions achieve 89.8% environmental (0.632), 97.7% social (0.751), and 79.5% economic (0.494) performance relative to single-objective maxima—demonstrating simultaneous optimization requires only modest compromises. The leading balanced solution (Trial 1, Solution 5) emerges at 16,197 persons/km², with top-performing balanced configurations clustering around 13,000–18,000 persons/km²—a range suggesting natural equilibrium where objectives reinforce rather than conflict.

The top-performing balanced solutions consistently appear in census tracts adjacent to or containing significant park space. Balanced solutions converge on 25.3% open space through emergent optimization, not design constraint (Supplementary Table 3, Fig. 4). This represents a quantitative tipping point where parks generate neighborhood-wide benefits while maintaining economic viability. Despite using only 65% as much open space as environmental optimizers (25.3% vs 38.8%), balanced configurations achieve 89.8% of maximum environmental performance. Strategic placement matters more than total quantity. Parks embedded in mixed-use, transit-accessible neighborhoods at moderate densities generate more total value than low-density preserves.

The optimization identified a density threshold: balanced solutions cluster at 13,000–18,000 persons/km², with the leading solution at 16,197 persons/km². Environmental leaders maintain densities below 4,881 persons/km², while social leaders span 15,301 − 33,445 persons/km². Notably, the highest observed density achieving Pareto optimality is 33,445 persons/km², suggesting diminishing sustainability returns beyond this threshold. The balanced land use mix—25.3% open space, 10.9% low, 8.7% medium, 55.1% high-intensity—represents careful calibration where each type contributes to overall performance. Graduated intensity levels create transition zones preserving neighborhood character while accommodating growth. These configurations provide immediately actionable templates: planners can reference specific land use mixes proven mathematically optimal for different policy priorities.

By relying on freely available census data, satellite imagery, and open-source machine learning libraries, this framework democratizes sophisticated multi-objective optimization previously limited to wealthy cities with extensive technical capacity¹⁴. Any urban area with basic census statistics and environmental data can apply this approach enabling comprehensive city-wide analysis. This accessibility addresses power asymmetries in the emerging urban cognitive ecosystem, where AI capabilities have concentrated in wealthy municipalities. By using only open data sources and reproducible methods, we distribute cognitive infrastructure as readily as we advocate for physical green infrastructure. This accessibility proves critical as cities generate approximately 70% of global carbon emissions¹⁹, and face intensifying climate impacts from extreme heat, flooding, and storms²⁰, which makes urban land use optimization essential for achieving 2030 and 2050 climate targets. This computational approach addresses the "planning gap" where cities lack tools to operationalize sustainability commitments⁵. The framework's iterative nature enables cities to continuously update optimization as census data, climate projections, or policy priorities evolve, supporting adaptive planning that responds to climate impacts, demographic shifts, and infrastructure changes, addressing the need for dynamic, evidence-based urban planning identified in SDG 11.3 on inclusive and sustainable urbanization¹⁸. This framework also establishes foundations for integrated systems modeling. Coupling land use optimization with energy and transportation models would enable comprehensive decarbonization strategies, moving from static pattern optimization to dynamic transition pathways that co-optimize across urban systems to achieve climate targets while maintaining equity and economic vitality.

Several limitations warrant acknowledgment. Optimization captures a sustainability snapshot but misses temporal dynamics including climate change impacts and development trajectories. Census tract-level analysis cannot capture within-neighborhood inequities—optimization may improve averages while exacerbating disparities for vulnerable subpopulations. The three-objective formulation, while tractable, may oversimplify urban complexity. Future work should develop dynamic optimization incorporating climate projections, integrate citizen preferences through participatory modeling, and conduct multi-city comparative analysis with finer granularity data to distinguish universal patterns from context-specific findings. Despite these limitations, this framework provides a scalable, open-source foundation enabling evidence-based urban sustainability planning previously accessible only to wealthy cities, bridging computational efficiency with participatory governance.