Frequency-aware vision transformers for high-fidelity super-resolution of Earth system models.

Area of Science:

• Climate Science
• Earth System Modeling
• Artificial Intelligence

Background:

• Coarse Earth System Model (ESM) simulations lack spatial fidelity crucial for climate science.
• Traditional deep super-resolution methods often introduce spectral bias, favoring low-frequency details over high-frequency information.

Purpose of the Study:

• To develop novel frequency-aware deep learning frameworks to mitigate spectral bias in ESM super-resolution.
• To improve the reconstruction of fine-scale structures and high-frequency details in climate data.

Main Methods:

• Introduced ViSIR (Vision Transformer-Tuned Sinusoidal Implicit Representation) combining vision transformers with sinusoidal activations.
• Developed ViFOR (Vision Transformer Fourier Representation Network) integrating explicit Fourier filtering for independent frequency learning.
• Evaluated models on the E3SM-HR Earth system dataset for surface temperature, shortwave, and longwave fluxes.

Main Results:

• ViSIR and ViFOR outperformed Convolutional Neural Networks (CNNs), Generative Networks, and vanilla transformers.
• ViFOR achieved up to 2.6 dB improvement in Peak Signal to Noise Ratio (PSNR).
• Both models demonstrated superior Structural Similarity (SSIM) compared to baseline methods.

Conclusions:

• Frequency-aware frameworks like ViSIR and ViFOR effectively address spectral bias in ESM super-resolution.
• These models significantly enhance the spatial fidelity of climate simulation outputs.
• ViFOR shows particular promise for high-fidelity climate data reconstruction.