In an era increasingly shaped by climate variability and its far-reaching impacts, the ability to forecast short-term sea level variations holds critical importance for coastal resilience and maritime operations. Temporary shifts in sea level—driven by changes in atmospheric pressure, wind patterns, and storm-induced water displacement—pose a significant threat to coastal communities through flooding and disrupt shipping, fisheries, and offshore infrastructure. Addressing these challenges requires innovative forecasting approaches that provide timely, precise predictions, enabling preemptive measures and efficient resource allocation.
Sea level anomaly (SLA) emerges as a pivotal metric in this realm, representing deviations in sea surface height relative to a long-term average. SLA embodies the fluctuations of geostrophic ocean currents, denoting discrepancies between present ocean surface states and their climatological baseline. This measure, rendered accessible through satellite altimetry, offers a global vantage point of sea surface dynamics, capturing subtle but consequential changes that precede coastal impacts. However, translating these data into accurate forecasts remains a formidable task.
Numerical models have traditionally driven short-term SLA prediction efforts. By integrating satellite altimetry data and physical oceanographic principles, these models simulate ocean circulation and sea level dynamics to forecast SLA trends. Yet, numerical approaches often grapple with persistent biases, limited resolution, and high computational demands that restrict their applicability—and accuracy—particularly in regional or near-coast contexts where high precision is paramount. The quest for improvement has catalyzed a shift toward data-driven methodologies.
Artificial intelligence (AI) has revolutionized many scientific domains by leveraging vast datasets to unveil complex, nonlinear patterns—capabilities that classical numerical models occasionally struggle to capture. In marine sciences, AI-driven ocean forecasting systems have begun to surpass traditional approaches in skill, especially across 10-day prediction horizons. These large-scale Global Ocean Forecast Systems (GOFSs), while effective, are tailored for global operations and require substantial computational resources, impeding their deployment for localized, resource-limited applications.
Recognizing these constraints, a cross-institutional team of researchers from Sun Yat-Sen University, Zhejiang Institute of Marine Planning and Design, and Pusan National University pursued an alternative route. Their aim was to enhance regional sea level prediction capacity in the North Pacific by refining AI training protocols without escalating the complexity of model architectures. This strategic pivot underscores a paradigm shift from model intricacy to model training efficacy.
The researchers’ work was recently published in Ocean-Land-Atmosphere Research, framing a novel approach centered on optimizing the training methodologies of AI models for SLA forecasting. Rather than constructing more intricate neural networks, they revisited the assumptions underlying prediction targets and temporal training frameworks to diminish errors that typically accumulate over extended forecasting intervals. The emphasis on improved training decoration rather than architectural modification holds promise for broad applicability.
One breakthrough involved redefining the forecasting focus from absolute SLA values toward the temporal tendency of SLA—that is, the day-to-day change in sea level anomalies. This subtle reframing captures slower-evolving dynamics more effectively, aligning prediction goals with the physical behavior of ocean currents and mitigating noise introduced by rapid, smaller-scale fluctuations. Another innovation dealt with the critical challenge of the “training-forecast gap.”
The “training-forecast gap” refers to the mismatch arising when models are trained on short-term forecast horizons yet deployed for longer-range predictions, leading to error propagation and reduced reliability over time. The team addressed this by employing a multi-step training paradigm and rolling forecast technique. This strategy entailed training the model primarily for daily SLA tendencies but enabling it to generate reliable forecasts over broader horizons through sequential, stepwise updates—akin to effectively chaining short-term predictions into robust medium-range outlooks.
At the heart of the AI architecture lies Earthformer, a cutting-edge deep learning model designed to process spatiotemporal data in parallel rather than sequentially. This parallelism enables efficient handling of complex oceanographic datasets while capturing crucial temporal correlations. Tailoring Earthformer to the extrinsic characteristics of the North Pacific’s altimetry data and imbuing it with optimized training schemata, the researchers crafted the Multistep-Earthformer forecasting system.
Comparative evaluations underscored the superiority of this approach. The Multistep-Earthformer significantly outperformed conventional approaches, including persistence forecasts—which naively assume future states will mirror the present—and even benchmarked numerical models such as GLO12v4. These improvements did not necessitate complicated model modifications but rather stemmed from intelligent training adjustments, suggesting a scalable pathway for other geophysical forecasting challenges.
Beyond sea level prediction, the conceptual advancements introduced—the focus on temporal tendencies and training regimen refinement—may invigorate broader applications in geosciences. AI modelers working on atmospheric, terrestrial, or coupled Earth system predictions could potentially adapt these strategies to mitigate forecast drift and error accumulation, thereby elevating forecast reliability across domains.
With the promising results in the North Pacific region as a foundation, the research collective now envisages scaling their approach globally. This expansion will demand careful tailoring of training strategies to accommodate regional oceanographic peculiarities and data availability. Their ultimate vision is the realization of an AI-driven, globally comprehensive sea level forecasting platform capable of supporting diverse stakeholders, from environmental managers to maritime industries.
This breakthrough emerges in the context of a rapidly evolving scientific landscape where interdisciplinary collaboration—spanning oceanography, AI, and computational modeling—fuels progress. The study received support from the Guangdong Basic and Applied Basic Research Foundation and Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), exemplifying the vital role of sustained funding and institutional synergy in addressing pressing environmental challenges.
Co-authors include Yong Liu of the Zhejiang Institute of Hydraulics & Estuary, Guangyu Yang of Sun Yat-Sen University, Young-Heon Jo of Pusan National University, and Zhigang Lai of Sun Yat-Sen University, reflecting a collaborative network bridging China and South Korea in marine sciences innovation.
As coastal regions worldwide confront intensifying climate-related hazards, tools that enable anticipatory action through advanced sea level forecasting will be indispensable. The Multistep-Earthformer model and its training-focused innovations provide a beacon of progress, illustrating how methodical refinement rather than mere complexity can unlock new horizons in ocean state prediction. This blend of sophisticated AI design and pragmatic training strategy exemplifies the next wave of environmental modeling poised to better safeguard coastal futures.
Subject of Research: Not applicable
Article Title: Optimized Training Strategies for AI-Based Sea Level Anomaly Forecasting in the North Pacific Ocean
News Publication Date: 23-Jan-2026
Web References: 10.34133/olar.0128
Image Credits: Jiangnan He et al. / Ocean-Land-Atmosphere Research
Keywords: Ocean circulation, Ocean physics, Oceanography, Air-sea interactions
