Reconstructing the Tropical Pacific Upper Ocean Using Online Data Assimilation With a Deep Learning Model.

Area of Science:

• Climate Science
• Machine Learning
• Oceanography

Background:

• Traditional linear inverse models (LIM) are standard for climate forecasting.
• Deep learning (DL) models offer potential for improved accuracy in complex climate systems.

Purpose of the Study:

• To compare the forecasting accuracy of a transformer-based DL model against a LIM in the tropical Pacific.
• To evaluate the DL model's effectiveness in reconstructing upper ocean conditions using simulated coral proxy data.

Main Methods:

• Training a DL transformer model on climate data.
• Comparing DL model forecasts with LIM forecasts using reanalysis data.
• Assessing ocean reconstruction using an ensemble Kalman filter with simulated sea-surface temperature observations.
• Implementing a novel noise inflation technique for the DL model.

Main Results:

• The DL model demonstrated higher forecast accuracy than the LIM on reanalysis data.
• DL model-based data assimilation yielded superior ocean reconstructions compared to LIM.
• Improved reconstructions were observed across various observation averaging times (1 month to 1 year).

Conclusions:

• Deep learning models, particularly transformer architectures, show significant promise for enhancing climate forecasting and ocean state reconstruction.
• The DL model's ability to map past observational memory to future assimilation times is key to its improved predictive performance.
• Novel techniques like noise inflation can address challenges such as signal damping in DL models for climate applications.