Improved Forecasts of Global Extreme Marine Heatwaves Through a Physics-guided Data-driven Approach

TL;DR

Facing the challenge of predicting extreme marine heatwaves (MHWs), this work develops a physics-guided data-driven framework that combines a coupler for atmospheric forcing with a probabilistic data-augmentation pathway to emphasize extremes, enabling 10-day global forecasts. The deterministic component models heat flux and advection/mixing via two neural nets, while a VQVAE-based ensemble-forecast approach generates diverse, physically plausible extremal states for retraining. The approach yields large gains over prior data-driven methods and competitive accuracy with numerical models, while delivering exceptional computational efficiency. Explainability analyses identify wind forcing and latent heat flux as the primary drivers of MHW evolution, providing both physical insight and a basis for operational forecasting and further research. The work demonstrates a scalable path to improved extreme-event forecasting and motivates incorporating subsurface data and refined atmospheric forcing in future iterations.

Abstract

The unusually warm sea surface temperature events known as marine heatwaves (MHWs) have a profound impact on marine ecosystems. Accurate prediction of extreme MHWs has significant scientific and financial worth. However, existing methods still have certain limitations, especially in the most extreme MHWs. In this study, to address these issues, based on the physical nature of MHWs, we created a novel deep learning neural network that is capable of accurate 10-day MHW forecasting. Our framework significantly improves the forecast ability of extreme MHWs through two specially designed modules inspired by numerical models: a coupler and a probabilistic data argumentation. The coupler simulates the driving effect of atmosphere on MHWs while the probabilistic data argumentation approaches significantly boost the forecast ability of extreme MHWs based on the idea of ensemble forecast. Compared with traditional numerical prediction, our framework has significantly higher accuracy and requires fewer computational resources. What's more, explainable AI methods show that wind forcing is the primary driver of MHW evolution and reveal its relation with air-sea heat exchange. Overall, our model provides a framework for understanding MHWs' driving processes and operational forecasts in the future.

Figures (5)

• Figure 1: The overall framework of our global MHWs forecast framework. a) The deterministic forecast. b) The probabilistic part designed to improve extreme MHWs' prediction.
• Figure 2: The overall performance of our global MHWs forecast model. a)-c) The RMSE, CSI, SEDI of our framework comparing to existing subseasonal MHWs forecast at a 10-day lead. d) The CSI score around global coastal regions. e) The 10-day roll-out forecast of our framework initialized on 2021-08-22. f) The forecast result (SSTA) of our framework initialized on 2021-10-03. It is important to highlight that we deduct 90th percentile of the original SSTA data, hence the red part of this image indicates the MHW event (if it also lasts longer than 5 days).
• Figure 3: A comparison of our deterministic and probabilistic forecasts. a)-b) The RMSE and CSI of our deterministic and probabilistic forecasts. c) The probability density function (PDF) of our deterministic and probabilistic forecast results. d)-e) A 10-day forecast of 2020 CNP-MHW initialized on 2020-08-14.
• Figure 4: The performance of our model in the high-resolution regional forecast (North Pacific). a) The 10-day forecast result of MHWs on 2020-07-17. The purple to orange parts of the image represent MHW events and the shade of color represents its intensity. b)-c) A comparison of our deterministic and probabilistic forecasts, similiar to Figure 3a-b. d) The power spectral of SSTA forecast at different lead time.
• Figure 5: The explainability of our model. a) The contribution map of each input variables on the SSTA during intensifying/decaying phase of MHWs. From top to the bottom: unsigned map during intensifying phase and decaying phase, signed map during intensifying and decaying phase. b) The unsigned contribution map of each input variables on the coupler's outputs during the intensifying phase of MHWs.