Ocean-E2E: Hybrid Physics-Based and Data-Driven Global Forecasting of Extreme Marine Heatwaves with End-to-End Neural Assimilation

Ruiqi Shu, Yuan Gao, Hao Wu, Ruijian Gou, Yanfei Xiang, Fan Xu, Qingsong Wen, Xian Wu, Xiaomeng Huang · 2025

Read Ocean-E2E: Hybrid Physics-Based and Data-Driven Global Forecasting of Extreme Marine Heatwaves with End-to-End Neural Assimilation by Ruiqi Shu, Yuan Gao, Hao Wu, Ruijian Gou, Yanfei Xiang, Fan Xu, Qingsong Wen, Xian Wu, Xiaomeng Huang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-22071.txt.

Introduction

Ruiqi Shu and co-authors present a forecast system for extreme ocean heat in this 2025 paper. Marine heatwaves are spells of unusually warm sea surface that can harm sea life badly. Forecasting the extreme ones has clear value, but past methods handle broad patterns and extremes poorly.

Their system, called Ocean-E2E, mixes physics-based ideas with data-driven learning. It models how ocean swirls move heat and how sea and air affect each other, using a changing kernel. It can run a full forecast on its own and give fine detail for single regions. The authors report 40-day forecasts that beat leading ocean models, with the best gains on the most extreme events.

Catalog record

CREATORRuiqi Shu, Yuan Gao, Hao Wu, Ruijian Gou, Yanfei Xiang, Fan Xu, Qingsong Wen, Xian Wu, Xiaomeng Huang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-22071

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Ocean-E2E: Hybrid Physics-Based and Data-Driven Global Forecasting of Extreme Marine Heatwaves with End-to-End Neural Assimilation Ruiqi Shu, Yuan Gao, Hao Wu, Ruijian Gou, Yanfei Xiang, Fan Xu, Qingsong Wen, Xian Wu, Xiaomeng Huang Abstract This work focuses on the end-to-end forecast of global extreme marine heatwaves (MHWs), which are unusually warm sea surface temperature events with profound impacts 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 hybrid data-driven and numerical MHWs forecast framework Ocean-E2E, which is capable of 40-day accurate MHW forecasting with end-to-end data assimilation. Our framework significantly improves the forecast ability of extreme MHWs by explicitly modeling the effect of oceanic mesoscale advection and air-sea interaction based on a differentiable dynamic kernel. Furthermore, Ocean-E2E is capable of end-to-end MHWs forecast and regional high-resolution prediction using neural data assimilation approaches, allowing our framework to operate completely independently of numerical models while demonstrating high assimilation stability and accuracy, outperforming the current state-of-the-art ocean numerical forecasting-assimilation models. Experimental results show that the proposed framework performs excellently on global-to-regional scales and short-to-long-term forecasts, especially in those most extreme MHWs. Overall, our model provides a framework for forecasting and understanding MHWs and other climate extremes. Our codes are available at https://github.com/ChiyodaMomo01/Ocean-E2E . 1 Introduction Marine heatwaves (MHWs) are abnormally warm seawater events that significantly damage marine ecosystems oliver2021marine ; pearce2011marine . In other words, MHWs are extreme Sea Surface Temperature Anomaly (SSTA) events. MHWs, particularly extreme ones, can cause coral bleaching hughes2017global ; hughes2018spatial and widespread mortality of marine organisms garrabou2009mass ; thomson2015extreme . As a result, precise forecasting of extreme MHWs has significant scientific and economic implications. For example, synoptic scale MHWs forecasting can help sea food production and management planning, such as feed cycles, at 1-7 day timescales, while subseasonal to seasonal forecasting can further support proactive decision-making for the blue economy hobday2016seasonal ; malick2020environmentally ; mills2017forecasting ; payne2022skilful . In this study, we will mainly focus on the subseasonal-to-seasonal forecast (i.e., 1-40 day). Traditional MHW forecasting predominantly relies on numerical models solving oceanic primitive equations, which generally bifurcate into two categories: seasonal forecasts characterizing large-scale patterns of MHW onset and intensification jacox2022global ; brodie2023ecological , and sub-seasonal predictions resolving finer-scale variability with reduced lead times benthuysen2021subseasonal ; yu2024assessing . While these physics-based approaches achieve operational utility, they demand substantial computational resources xiong2023ai ; wang2024xihe and exhibit limited skill in predicting extreme MHW events jacox2022global . Recent advances in deep learning have demonstrated potential for efficient global ocean forecasting bi2022pangu ; chen2023fengwu ; kurth2023fourcastnet , with emerging applications to MHW prediction lin2023rossby ; sun2023artificial . However, existing implementations face dual limitations: 1) insufficient encoding of MHW-specific physics —particularly air-sea coupling mechanisms driving extreme events and mesoscale eddy dynamics governing local intensification—resulting in non-physical smoothing of SST gradients during autoregressive prediction giamalaki2022assessing ; 2) dependence on numerical model initialization that prevent operational end-to-end forecasting from raw observational inputs xiong2023ai . These architectural constraints—shared across mainstream deep learning paradigms—fundamentally limit their capacity to preserve extreme SST anomalies and long-term physical consistency , as error accumulation from neural networks and initialization biases during long-term rollouts wang2024xihe . By addressing the two shortcomings of the data-driven methods mentioned above, in this study, we designed Ocean-E2E, a global data-driven framework that significantly improves the forecasts of extreme MHWs inspired by numerical models. The contribution of this paper can be summarized as follows: (1) Global MHWs Forecasting Framework . We propose a hybrid physics-based data-driven method that supports both global scale and regional high-resolution forecasting, achieving high-accuracy results for extreme MHWs forecasts. (2) Global MHWs Neural Assmilation Framework . We designed an end-to-end MHWs neural system that, when combined with the forecasting model, can directly obtain the complete initial field and forecasting results of MHWs from observation data. (3) Long-Term Forecast . By combining phyiscal laws into our framework, our model has achieved state-of-the-art results in long-term (subseasonal-to-seasonal) forecasting with higher physical consistency. The method most similar to us is ClimODE verma2024climode . Similarly, it also uses advection partial differential equations to design the forecasting framework. However, our work has many differences and improvements compared to ClimODE: (1) Differences in theoretical framework. From the perspective of the theoretical framework, in their work, the ‘speed’ of the flow field is essentially an abstract mathematical definition, whose core function is to maintain the conservation of passive tracers. In our work, the flow field speed strictly corresponds to the realistic ocean flow field and models the additional transport effect of the flow field on passive tracers at sub-grid scales through mesoscale ‘bolus’ velocity. (2) Differences in model design. ClimODE uses an emission model to simulate the unclosed and uncertain portions of the real system, while our model further relates this part to the ocean-atmosphere interaction and is designed based on physical theory to form the structure of the source term. (3) Differences in application. ClimODE requires solving the initial velocity field as the initial condition for the model, which involves solving a physics-informed neural network (PINNs) for each training sample. This makes the ClimODE framework difficult to scale up. What’s more, our experiments (Section 4.3 ) show that our method achieves greater stability and accuracy in MHWs forecasting compared to ClimODE. Figure 1: Our proposed Ocean-E2E framwork. a) Overview of our framework. b) Details of our hybrid physics and data-driven forecast model. 2 Problem Setup In this study, the problem of the end-to-end MHWs forecast is essentially forecasting SSTA , which can be separated into two parts. First, estimating the current state of SSTA T_{0}\in\mathbb{R}^{H\times W} (a global two-dimensional field, where H and W correspond to latitude and longitude) based on the observation T_{0}^{obs} and the background field T_{0}^{bg} . Second, predicting the future state of mixed layer SSTA T_{t} based on the current state T_{0} under atmospheric forcing A_{0:t}\in\mathbb{R}^{T\times H\times W} . Here the underscore 0:t denotes the atmosphere state from time 0 to t . Formally speaking, we are going to model the conditional distribution p(T_{t}|T_{0}^{obs},T_{0}^{bg},A_{0:t},) in two stages: p(T_{t}|T_{0}^{obs},T_{0}^{bg},A_{0:t})=\underbrace{p^{f}(T_{t}|T_{0},A_{0:t})}_{\text{Forecast model}}\cdot\underbrace{p^{a}(T_{0}|T_{0}^{obs},T_{0}^{bg})}_{\text{Assimilation model}} (1) where p^{f} and p^{a} denote forecast and assimilation process, respectively. In our end-to-end framework, we model these two processes through two specifically designed neural network \phi^{f}_{\theta} and \phi^{a}_{\theta} : T_{t}=\phi^{f}_{\theta}(T_{0},A_{0:t}) (2) T_{0}=\phi^{a}_{\theta}(T_{0}^{obs},T_{0}^{bg}) (3) In summary, our end-to-end framework aims to forecast the furture state of MHWs T_{t} based on the sparse observation T_{0}^{obs} and the initial background field T_{0}^{bg} under the forcing of atmosphere boundary conditions A_{0:t} . The parameters of these two neural networks \theta_{f},\theta_{a} are optimized separately by minizing the mean square error (MSE) between the model’s output and the groundtruth: \theta_{f}={\rm argmin}\{\frac{1}{N}\sum_{i=0}^{N}(T_{t}^{(i)}-\phi^{f}_{\theta_{f}}(T_{0}^{(i)},A_{0:t}^{(i)}))^{2}\} (4) \theta_{a}={\rm argmin}\{\frac{1}{N}\sum_{i=0}^{N}(T_{t}^{(i)}-\phi^{a}_{\theta_{a}}(T_{0}^{(i)},A_{0:t}^{(i)}))^{2}\} (5) 3 Methods 3.1 Motivation and Framework Overview Our framework is summarized in Figure 1 . As established in the problem formulation, forecasting MHWs fundamentally reduces to predicting extreme sea surface temperature anomalies (SSTA). To effectively embed physical principles into our modeling framework, we must rigorously account for the governing laws of SSTA evolution. We approach this challenge through a generalized geophysical fluid dynamics perspective, wherein seawater temperature constitutes a

Read the complete text