EN / TR
← Back to Research Lab
06 — DISTRICT Ongoing 2024-2025

Computational Urbanism

Algorithmic Form-Finding at District Scale

Computational Urbanism Research
Urban density optimization study for London metropolitan area

How do you design a neighborhood when there are 10^15 possible configurations of building placement, street width, and block size? Traditional masterplanning relies on intuition and precedent. This research takes a different approach: we let algorithms explore the design space.

Using evolutionary algorithms (NSGA-II), agent-based pedestrian simulation (127,000 agents), and multi-objective optimization, we generate thousands of urban configurations and filter for solutions that satisfy conflicting objectives: maximize density, maximize walkability, maximize solar access, minimize wind discomfort.

The key finding: optimal solutions are non-obvious. When we optimized for both density AND walkability on the Bilkent campus site (50 ha), the algorithm consistently produced polycentric clusters rather than the uniform distribution that intuition suggests—a result that aligns with Jane Jacobs' observations but was discovered computationally.

Research Process

01

Data Acquisition

GIS footprints from OpenStreetMap (12,847 buildings within 2km radius), traffic flow from Uber Movement (4 years), demographic patterns from TÜİK census API. All data projected to EPSG:32636.

Python QGIS GeoPandas
12,847 buildings mapped
02

Agent-Based Modeling

127,000 pedestrian agents with heterogeneous behaviors (commuters, shoppers, students, elderly) navigate the urban fabric over simulated 24-hour cycles. Agents record path choice, dwell time, and congestion events. Calibrated against manual pedestrian counts (r² = 0.89).

Python Mesa ABM Grasshopper
127k agents × 24hr cycles
03

Evolutionary Optimization

NSGA-II with population size 100, 50 generations (5,000 total evaluations). Four fitness objectives: FAR (maximize), walkability score (maximize), solar access (maximize winter, minimize summer), wind comfort at pedestrian level (<5 m/s). Pareto front yields 23 non-dominated solutions.

Wallacei Grasshopper Galapagos
5,000 designs evaluated
04

Environmental Simulation

Each of the 5,000 designs receives a 1-minute computational budget: solar radiation (Ladybug, 12 hours), wind CFD (Butterfly, simplified boundary conditions), thermal comfort (UTCI). Total compute: 83 GPU-hours on NVIDIA A5000.

Ladybug Honeybee Butterfly CFD
83 GPU-hours total
Pedestrian flow heatmap simulation

127,000 Agent Simulation: Pedestrian flow heatmap showing congestion hotspots and movement corridors. Origin-destination analysis reveals that small block patterns (<100m perimeter) naturally optimize for walkability.

What We Discovered

01

Evolutionary algorithms explore 10,000+ design iterations in 24 hours—a search space impossible for human designers to navigate manually

10,000+ iterations / 24hr
02

NSGA-II reveals non-obvious trade-offs: maximizing both density AND walkability requires polycentric clustering (5 micro-centers), not uniform distribution

5 clusters optimal
03

Agent-based pedestrian simulation predicts congestion points with 87% accuracy compared to post-occupancy evaluation on a reference site (Kadıköy)

87% prediction accuracy
04

Small block sizes (<100m perimeter) emerge naturally from swarm optimization when walkability is weighted ≥0.3 in the fitness function

<100m block size

Current Limitations

Data dependency: model accuracy depends heavily on input quality; missing or outdated GIS data produces unreliable predictions

Computational cost: full-fidelity CFD for wind comfort is prohibitive at scale; we use simplified boundary conditions that may miss microclimate effects

Behavioral assumptions: agent parameters (walking speed, destination choice) are calibrated to Turkish cities and may not transfer to other contexts

No temporal dynamics: the model optimizes for a static 'target year' and does not account for phased development or changing demographics

Interested in Urban Research?

We're always looking for research partners, municipalities, and developers interested in data-driven urbanism.