Abstract:Machine learning (ML) methods have shown great potential for weather downscaling. These data-driven approaches provide a more efficient alternative for producing high-resolution weather datasets and forecasts compared to physics-based numerical simulations. Neural operators, which learn solution operators for a family of partial differential equations (PDEs), have shown great success in scientific ML applications involving physics-driven datasets. Neural operators are grid-resolution-invariant and are often evaluated on higher grid resolutions than they are trained on, i.e., zero-shot super-resolution. Given their promising zero-shot super-resolution performance on dynamical systems emulation, we present a critical investigation of their zero-shot weather downscaling capabilities, which is when models are tasked with producing high-resolution outputs using higher upsampling factors than are seen during training. To this end, we create two realistic downscaling experiments with challenging upsampling factors (e.g., 8x and 15x) across data from different simulations: the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) and the Wind Integration National Dataset Toolkit (WTK). While neural operator-based downscaling models perform better than interpolation and a simple convolutional baseline, we show the surprising performance of an approach that combines a powerful transformer-based model with parameter-free interpolation at zero-shot weather downscaling. We find that this Swin-Transformer-based approach mostly outperforms models with neural operator layers, and suggest its use in future work as a strong baseline.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve The paper aims to explore the effectiveness of Neural Operators in zero-shot weather downscaling. Specifically, it investigates whether Neural Operators can use a lower upsampling factor (e.g., 4x) during training and generate high-resolution outputs with higher upsampling factors (e.g., 8x or 15x) during testing. This capability is significant for improving the accuracy of weather forecasting and climate research. ### Background and Motivation 1. **Importance of Weather Downscaling**: - Weather downscaling techniques are used to obtain high-resolution (HR) data from low-resolution (LR) data. These high-resolution data can provide more detailed information on physical phenomena, aiding in the analysis of extreme weather events, optimizing grid planning, and managing renewable energy (e.g., wind energy). 2. **Limitations of Traditional Methods**: - Physically-driven numerical simulations, while accurate, are computationally expensive in practical applications. Data-driven methods, especially deep learning methods, have gained attention for their efficiency. 3. **Advantages of Neural Operators**: - Neural Operators can learn solution operators for a class of partial differential equations (PDEs) and are invariant to grid resolution. They can use low-resolution data during training and generate high-resolution outputs during testing, achieving zero-shot super-resolution. ### Experimental Design 1. **Experimental Setup**: - The authors designed two challenging downscaling experiments using data from the European Centre for Medium-Range Weather Forecasts Reanalysis 5th Edition (ERA5) and the Wind Toolkit (WTK). The experiments included 8x and 15x upsampling factors. 2. **Model Comparison**: - The authors compared various Neural Operator models (e.g., FNO, DFNO, DUNO, DCNO, DAFNO) and other baseline models (e.g., Bicubic Interpolation, SRCNN, EDSR, SwinIR) in standard downscaling and zero-shot downscaling tasks. ### Main Findings 1. **Performance Comparison**: - In standard downscaling tasks, SwinIR outperformed other models on all metrics (MSE, MAE, IN, PSNR), followed by DCNO. - In zero-shot downscaling tasks, SwinIR still performed the best, with DUNO being the best among Neural Operator models, but DCNO's performance significantly declined. 2. **Energy Spectrum Analysis**: - Through energy spectrum analysis, SwinIR's energy distribution was closest to the true high-resolution data in the low to mid wavenumber range, while DAFNO performed best in the highest wavenumber range. ### Conclusion 1. **Main Contributions**: - Provided a comparative analysis based on two challenging weather downscaling problems involving large upsampling factors (e.g., 8x and 15x) and fine grid resolutions (e.g., 2 km x 2 km wind speed). - Explored the ability of Neural Operator layers to downscale using higher upsampling factors during testing than during training. The results showed that combining powerful transformer models with non-parametric interpolation steps (e.g., SwinIR) performed excellently in zero-shot weather downscaling. - Suggested SwinIR as a strong baseline model for future work, but further improvements are needed to better capture small-scale physical characteristics, such as turbulence. 2. **Future Work Directions**: - Further optimize transformer and Neural Operator models to enhance their performance in zero-shot weather downscaling, particularly in capturing small-scale physical characteristics.

On the Effectiveness of Neural Operators at Zero-Shot Weather Downscaling

Fourier Neural Operators for Arbitrary Resolution Climate Data Downscaling

Exploring super-resolution spatial downscaling of several meteorological variables and potential applications for photovoltaic power

Downscaling of wind fields on the east coast of India using deep convolutional neural networks and their applications in storm surge computations

Interpolation-Free Deep Learning for Meteorological Downscaling on Unaligned Grids Across Multiple Domains with Application to Wind Power

Emerging AI-based weather prediction models as downscaling tools

A Novel Residual-Pyramid-attention Super Resolution Model for Mesoscale Meteorological Forecasting Spatial Downscaling

Deep learning in statistical downscaling for deriving high spatial resolution gridded meteorological data: A systematic review

Evaluating the transferability potential of deep learning models for climate downscaling

A comparative study of convolutional neural network models for wind field downscaling

Hard-Constrained Deep Learning for Climate Downscaling

A High-Resolution Land Surface Temperature Downscaling Method Based on Geographically Weighted Neural Network Regression

MetaSD: A Unified Framework for Scalable Downscaling of Meteorological Variables in Diverse Situations

Statistical treatment of convolutional neural network super-resolution of inland surface wind for subgrid-scale variability quantification

Wind Power Assessment based on Super-Resolution and Downscaling -- A Comparison of Deep Learning Methods

Efficient downscaling of satellite oceanographic data with convolutional neural networks

Deriving Accurate Surface Meteorological States at Arbitrary Locations Via Observation-Guided Continuous Neural Field Modeling.

Successful Precipitation Downscaling Through an Innovative Transformer-Based Model

Deep Learning for Daily 2‐m Temperature Downscaling

A Generative Deep Learning Approach to Stochastic Downscaling of Precipitation Forecasts

Scaling transformer neural networks for skillful and reliable medium-range weather forecasting