2024-SEBai

Overview

2024-SEBai is a machine learning based framework developed to bias-correct ERA5 surface fluxes over the Arctic Ocean using in situ observational measurements. The project applies a neural network model trained on observational datasets including MOSAiC, SHEBA, AO2018 and ARTofMELT field campaigns to improve the accuracy of ERA5 surface flux estimates, with a focus on Arctic conditions. Data from N-ICE2015 is used for testing, which is not used for training and from a different year/location than the training data. The aim is to verify that the bias correction model generalizes beyond the training/validation dataset.

The bias-corrected products are generated for the Arctic Ocean (north of 70°N) for the period 1994–2024.

Project Structure

Core Scripts

• datasets.py: Handles dataset loading and preprocessing.
• normalize.py: Performs normalization of the training datasets.
• neural_network.py: Defines the neural network architecture used for bias correction.
• utils.py: Contains utility functions for:
  • Model training (trainer)
  • Validation (validate)
  • Evaluation (evaluate_model, evaluate_model_arctic)
• training.ipynb: Jupyter Notebook used to train the neural network on all available datasets and to test the model using N-ICE data.
• bias_corrected_Arctic.py: Generates bias-corrected ERA5 surface fluxes over the Arctic Ocean (north of 70°N) for the period 1994–2024.

Data and Models

• data/: Contains all datasets required by the scripts, including observational data and ERA5 inputs.
• models/: Stores trained neural network models saved after the training process.

Usage

1. Model Training and Testing: Run training.ipynb to -
   • Train the neural network using the full training dataset
   • Test model performance using N-ICE observational data
2. Bias Correction of ERA5 Data Execute bias_corrected_Arctic.py to produce bias-corrected ERA5 surface fluxes for the Arctic region (north of 70°N) from 1994 to 2024.

Notes

• This project is designed specifically for Arctic surface flux correction and may require adaptation for other regions.
• Ensure all required datasets are available in the data/directory before running the scripts.