Big Data and Earth System 2025

Leveraging AutoML for Urban Climate Emulation

Efficient, low-carbon urban climate emulation via Automated Machine Learning (AutoML).

J. Yu, Z. Zheng, S. Lindley, L. Zhao, et al.

View Code Back to Home

Click anywhere to trigger particle explosion

Highlights

Automated Workflow: FLAML framework automatically optimizes hyperparameters without manual intervention.
Extreme Efficiency: Training speed increased.
High Accuracy: Achieved an R² > 0.98 for simulating urban climate variables.

Abstract

Democratizing Climate Emulation

Traditional Urban Climate Models (UCMs) are computationally expensive, making large-scale, long-term studies difficult. Existing machine learning emulators, while faster, often require tedious manual tuning.

This study proposes an Automated Machine Learning (AutoML) framework for urban climate emulation. Tested across cities with diverse climates (Beijing, Melbourne, Zurich), the framework demonstrates the ability to automatically generate high-performance emulators while drastically reducing computational resource usage. This provides a "plug-and-play" tool for urban planners and researchers.

Innovation

Fast & Lightweight

We adopt a lightweight AutoML strategy focused on finding the optimal model configuration under a minimal computational budget.

R² > 0.98

Exceptional fit for simulating both land surface temperature and atmospheric temperature.

Applications

Sustainable Planning

Rapidly assess the microclimatic impact of various urban design scenarios (e.g., increasing green spaces, changing building materials).

                    from flaml import AutoML

                    automl = AutoML()

                    automl.fit(X_train, y_train, task="regression")