End-to-end data-driven prediction of urban airflow and pollutant dispersion

Technical Deep Dive: End-to-end Data-Driven Prediction of Urban Airflow and Pollutant Dispersion

Problem & Context

• Climate change and urban population growth are intensifying environmental stresses within cities, making urban atmospheric flows critical for public health, energy use, and livability
• Computational fluid dynamics (CFD) models are commonly used to simulate urban airflow, but have high computational costs and are unsuitable for real-time forecasting or parametric studies
• This study aims to develop a data-driven reduced-order modeling (ROM) framework to efficiently predict urban pollutant dispersion, with a focus on the skimming flow regime in street canyons

Methodology

Data & Experimental Setup

• Large eddy simulation (LES) dataset of a 3D street canyon flow with a continuous pollutant source at the canyon base
• Snapshot data includes time-resolved streamwise and vertical velocity components, as well as scalar concentration
• Flow conforms to the skimming flow regime, characterized by a principal recirculation cell and smaller corner vortices

Reduced-Order Modeling

1. Spectral proper orthogonal decomposition (SPOD):
   • Extracts coherent spatial structures and their temporal evolution in the frequency domain
   • Dimensionality reduction via energy-based truncation and mode similarity criteria
2. Autoencoder (AE):
   • Learns a compact, nonlinear latent representation of the SPOD coefficients
3. Long short-term memory (LSTM):
   • Models the temporal evolution of the latent space variables
4. Convolutional neural network (CNN):
   • Maps the predicted velocity field to the corresponding pollutant concentration field

Results

SPOD Analysis

• SPOD effectively isolates key flow structures, such as the Kelvin-Helmholtz instability in the shear layer
• Dimensionality reduction retains 97% of the total turbulent kinetic energy using 2,003 SPOD modes

Latent Space Compression

• AE with 30-dimensional latent space captures the essential flow dynamics
• Reconstructed velocity fields exhibit good agreement with the reference LES data

Velocity Field Prediction

• LSTM accurately forecasts the temporal evolution of the latent variables, reproducing the phase space topology
• Velocity field reconstruction remains bounded and stable over long prediction horizons

Pollutant Dispersion Prediction

• CNN-based mapping from velocity to concentration field successfully captures the large-scale plume topology and ventilation characteristics
• Some smoothing of fine-scale structures, but overall good agreement with LES reference

Interpretation

• The proposed end-to-end data-driven ROM framework demonstrates the capability to efficiently predict both the velocity and pollutant concentration fields in an urban street canyon
• The modular architecture allows for flexible customization and future extensions, such as incorporating data assimilation or transfer learning
• The framework offers a computationally efficient alternative to high-fidelity CFD simulations, enabling real-time forecasting and extensive parametric studies for urban planning and air quality management

Limitations & Uncertainties

• Success of the framework relies on the ability of the SPOD modes to span the possible range of parameters; more configurations should be incorporated in the training dataset
• Training the neural networks is computationally expensive, but can potentially be addressed through techniques like transfer learning
• Validation is limited to a single street canyon configuration; further testing on diverse urban geometries is required to assess the generalization capabilities

What Comes Next

• Explore methods to incorporate additional physical constraints and domain knowledge into the neural network architectures
• Investigate strategies for online model adaptation and data assimilation to improve long-term prediction accuracy
• Extend the framework to handle time-varying boundary conditions and unsteady scenarios