The ClimateSimulation_Downscaling_GAN is a Conditional Generative Adversarial Network (cGAN) built for super-resolution spatial downscaling in climate modeling. Global Climate Models (GCMs) produce coarse outputs (e.g., $32 \times 32$ pixel grids). This model takes a coarse GCM output map (e.g., surface temperature anomaly) and generates a realistic, high-resolution counterpart ($256 \times 256$ pixels), introducing physically plausible fine-scale details.
🧠 Model Architecture
The model is based on the Pix2Pix framework, a variant of the cGAN designed for image-to-image translation tasks.
Generator Network: A U-Net-like architecture takes the low-resolution GCM output as input and attempts to generate the high-resolution target image. The U-Net structure helps capture features at multiple scales.
Discriminator Network: A PatchGAN discriminator attempts to classify whether a given high-resolution image (either the real target or the generated image) is fake or real, conditioned on the low-resolution input.
Input: Low-resolution gridded data (e.g., $32 \times 32$ map of a climate variable).
Output: High-resolution gridded data ($256 \times 256$ map of the same climate variable).
Training Objective: A combined loss function consisting of a standard GAN loss (to make the output look realistic) and an $L1$ loss (to ensure pixel-level fidelity to the true, high-resolution data).
🎯 Intended Use
This model is a major tool for climate impact assessment and regional planning:
Regional Impact Studies: Providing high-resolution climate forecasts necessary for local studies like water resource management, agricultural planning, and localized flood modeling.
Model Benchmarking: Generating realistic fine-scale data from coarse model runs, saving substantial computational time compared to running physical Regional Climate Models (RCMs).
Data Visualization: Creating compelling, high-detail climate visualizations for public and policy communication.
⚠️ Limitations
Generative Uncertainty: As a generative model, the output is not a unique, deterministic solution. The fine-scale details introduced are statistically plausible but may not represent the single true physical state.
Physical Consistency: The model is trained to minimize pixel error and maximize realism. It may occasionally violate strict conservation laws or physical constraints that are maintained by explicit physical models.
Bias Correction: The model is trained on a specific GCM output. If the low-resolution GCM has a systematic bias, the downscaling process will propagate and potentially amplify that bias into the high-resolution output.
