Fine-Tuning NeuralGCM with New Variables

[YohanAbeysinghe]

I’m interested in fine-tuning this model with an additional variable and would like to understand the feasibility of adapting the existing code for this purpose.

I noticed,

• The current fine-tuning process in trained_models.ipynb updates approximately 1.4 million parameters for the 2.8 resolution model, which appears to cover the full set of model parameters.
• But the fine-tuning currently relies solely on an L1 loss, unlike the full training code, which may incorporate other loss functions.
• Integrating the new variable will likely require adjustments to the dynamic core.

Would the current fine-tuning implementation support this modification, or would it require significant code changes? Any guidance on how to approach this would be appreciated!