Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Learning Urban Climate Dynamics via Physics-Guided Urban Surface–Atmosphere Interactions
Authors: Jiyang Xia, Fenghua Ling, Zhenhui Jessie Li, Junjie Yu, Hongliang Zhang, David Topping, LEI BAI, Zhonghua Zheng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that incorporating domain and physical knowledge leads to significant improvements in emulation accuracy and generalizability under future urban climate scenarios. Further analysis reveals that learning shared correlations across cities enables the model to capture transferable urban surface atmosphere interaction patterns, resulting in improved accuracy in urban climate emulation. Finally, UCformer shows strong potential to fit real-world data: when fine-tuned with limited observational data, it achieves competitive performance in estimating urban heat fluxes compared to a physics-based model. 1 |
| Researcher Affiliation | Collaboration | 1 The University of Manchester 2 Shanghai AI Laboratory 3 Yunqi Academy of Engineering 4 Fudan University |
| Pseudocode | No | The paper describes the model architecture and components through text and diagrams (Figure 2) and mathematical equations (e.g., Eq. 1-12), but it does not present any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | 1The code and datasets of this work are available at https://github.com/envdes/code_UCformer. |
| Open Datasets | Yes | 1The code and datasets of this work are available at https://github.com/envdes/code_UCformer. |
| Dataset Splits | Yes | The training set consists of data spanning from 2020 to 2044, with 8 time steps per day, totaling 1,095,000 data points. The validation and test sets comprise data from 2045–2049 and 2050–2055, each containing 219,000 data points. We further use the data from 2070 to 2074 to assess the generalizability of our model |
| Hardware Specification | Yes | We implemented the model using PyTorch and finalized its configuration via hyperparameter tuning with Optuna [2] in 35 GPU hours (NVIDIA 4090). ... All models were tested with a batch size of 128 on an NVIDIA 5090 GPU. |
| Software Dependencies | No | The paper mentions software like PyTorch, Optuna, and FLAML but does not provide specific version numbers for these components, which is required for a 'Yes' answer based on the guidelines. |
| Experiment Setup | Yes | We implemented the model using PyTorch and finalized its configuration via hyperparameter tuning with Optuna [2] in 35 GPU hours (NVIDIA 4090). The model was trained using the Adam optimizer and Gaussian Error Linear Unit (GELU) activation function, with a batch size of 64, a learning rate set to 1e-5, a training epoch of 50, and a dropout rate of 0.1. Appendix B.3 lists the detailed configuration of models. |