TODO: Hyperparameter tuning for surrogate model

[gijsschlief]

Investigate, design, and execute hyperparameter tuning for the surrogate model.

• Identify relevant parameters (e.g., learning rate, model architecture, batch size, etc.)
• Experiment with different values to improve accuracy and generalization
• Document process, results, and best practices
• Integrate tuned parameters into training pipeline

Acceptance Criteria:

• Surrogate model has demonstrably improved performance based on selected metrics
• All steps and findings are documented in the repository
• Update any related scripts/configs

---

If additional context or links to prior experiments exist, please reference them here.

[gijsschlief]

The tuning should at least include:

• batch size [16, 256]
• weight decay [1e-6, 1e-2]
• neural network width [512, 1024, 2048]
• neural network depth (half convolutional, half linear in decoder structure) [4, 6, 8]
• dropout [0, 0.5]
• combined loss alpha (MSE, SSIM ratio) [0, 1]