Urbanflow-3K contains 3,000 two-dimensional urban flow simulations generated using the lattice-Boltzmann method by Hojin Lee, published in 2026. The data captures time-averaged velocity fields across three distinct Reynolds numbers for layouts featuring three to six buildings with randomized configurations. This collection was developed to provide a lower-cost alternative for training machine learning models before scaling to three-dimensional datasets.
Use Cases
- Predicting time-averaged velocity fields from randomized building coordinates and rotation angles
- Benchmarking ML architectures on their ability to capture wake formation and flow acceleration
- Pre-training models for transfer learning to more computationally expensive 3D urban flow datasets
Strengths
- 3,000 unique simulation records
- High geometric diversity with randomized building rotations from 0 to 90 degrees
- Consistent simulation parameters across three distinct Reynolds numbers
Limitations
- Restricted to 2D flow fields which omit vertical turbulence and shielding effects found in 3D space
- Synthetic data generated via CFD rather than empirical field observations or wind tunnel measurements
Provenance
- Source
- Harvard Dataverse
- Collection Method
- synthetic
- Freshness
- Last updated March 2026.