Work

DARK: Deep Learning-Assisted Regression Kriging Framework for High-Resolution Urban Microclimate Reconstruction from CFD-Derived Virtual Sparse Observations

Year: 2026

Type: article

Abstract: High-resolution urban wind and thermal fields are essential for urban climate analysis, ventilation assessment, and outdoor thermal comfort evaluation. However, generating such fields over large urban domains using computational fluid dynamics (CFD) remains computationally expensive. Regression Kriging (RK), as a classical interpolation-based method, is widely used for spatial field reconstruction, but its performance is constrained by the trade-off among spatial resolution, computational cost, and reconstruction accuracy. To address this limitation, this study proposed a Deep learning-Assisted Regression Kriging (DARK) framework for reconstructing high-resolution urban wind and temperatures fields. The proposed framework selected a working spatial resolution for RK-based inference by evaluating the effects of spatial resolution on reconstruction accuracy and computational efficiency. This RK setting was then combined with deep learning-based refinement to enable high-resolution reconstruction. Shapley additive explanations (SHAP) was also employed to quantify the importance of urban morphological features and identify informative reference points for CFD-derived virtual sparse observations. The proposed DARK framework was evaluated on a full urban domain at 4 m × 4 m resolution. The final reconstruction achieved RMSE values of 0.419 m/s, 0.445 m/s, and 0.472 °C for U, W, and T, respectively, compared with 0.626 m/s, 0.463 m/s, and 0.477 °C from standalone RK. In addition, direct RK reconstruction required approximately 182 s per timestep, whereas DARK completed the reconstruction in approximately 0.35 s per timestep, corresponding to a 520× speedup at inference stage. The results demonstrate the feasibility of DARK as a case-specific, computationally efficient workflow for high-resolution urban microclimate reconstruction under sparse-observation conditions.

Source: Sustainability

DOI: 10.3390/su182211487

Authors: Youmin Xu, Xu Han, Shaoxiang Qin, Liangzhu Leon Wang

Institutions: University of Kansas, University of Notre Dame, McGill University, Concordia University

Language: English

Topic: Wind and Air Flow Studies

Subfield: Environmental Engineering

Field: Environmental Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,896

OpenAlex ID: W7225771399

FWCI:

Cites:

Cited by:

Related to: