Physics-Informed Tensor-Train ConvLSTM for Volumetric Velocity Forecasting of the Loop Current

Area of Science:

• Oceanography
• Artificial Intelligence
• Geophysics

Background:

• Accurate forecasting of the Loop Current (LC) and its eddies in the Gulf of Mexico (GoM) is crucial for ecosystem understanding and disaster mitigation.
• Current numerical models struggle to resolve the complex multi-scale dynamics of the LC, limiting forecast accuracy.

Purpose of the Study:

• To develop a physics-informed deep learning model for accurate 3D geo-spatiotemporal forecasting of the LC and its eddies.
• To improve upon existing methods by incorporating physical laws into the deep learning framework.

Main Methods:

• Developed a Physics-informed Tensor-train ConvLSTM (PITT-ConvLSTM) model for 3D geo-spatiotemporal sequence prediction.
• Utilized a 4D higher-order recurrent neural network with empirical orthogonal function analysis and convolutional tensor-train decomposition.
• Integrated prior physics knowledge into the latent space learning process.

Main Results:

• The PITT-ConvLSTM model successfully forecasts the volumetric velocity of the LC and its eddies.
• Forecasts were achieved for a period exceeding one week.
• Demonstrated the model's ability to capture complex space-time correlations and frame dependencies.

Conclusions:

• Physics-informed deep learning offers a powerful approach for complex oceanographic forecasting.
• The PITT-ConvLSTM model provides a significant advancement in predicting the behavior of the Loop Current.
• This technology has implications for operational oceanography and risk management in the GoM.