Sea level rise (SLR) poses a significant threat to coastal communities globally, causing increased flooding, erosion, and saltwater intrusion. Accurate prediction of coastal sea levels is crucial for effective community preparedness. Sea level varies across a wide range of timescales, from hourly to centennial, with global SLR driven by thermal expansion and melting ice. Hourly to daily changes are mainly due to storm surges, tides, and waves, while seasonal to multidecadal variations are linked to climate variability (e.g., El Niño-Southern Oscillation) and large-scale ocean dynamics (e.g., AMOC). Multiyear-to-decadal variations form a baseline that modulates the impacts of higher-frequency events. The US East Coast is particularly vulnerable to SLR, with accelerated rise observed both north and south of Cape Hatteras. North of Cape Hatteras, the accelerated SLR is often attributed to sterodynamic effects and mass redistribution related to a weakened AMOC, leading to a sea level gradient across the Gulf Stream and North Atlantic Current. South of Cape Hatteras, the accelerated SLR since 2010 has multiple proposed explanations, including Gulf Stream variations, warming of the Florida Current, large-scale heat divergence linked to AMOC and NAO, a lagged response to AMOC slowdown, and wind-forced Rossby waves. The limited spatial and temporal coverage of tide gauges and satellite observations makes it challenging to fully capture North Atlantic sea level variability on decadal timescales. This study aims to overcome this limitation by employing a machine learning approach using long climate model simulations to better understand multiyear-to-decadal sea level variability and predictability in the North Atlantic.

Numerous studies have investigated the impacts of sea level rise on coastal communities, highlighting the urgent need for accurate sea level predictions. Research has established links between sea level variability and various climate phenomena, including ENSO and the AMOC. Studies have identified regional hotspots of accelerated SLR, particularly along the US East Coast, with different mechanisms proposed for regions north and south of Cape Hatteras. Existing studies have used various methods to investigate sea level predictability, but the limited length of observational records often hinders the isolation of low-frequency signals crucial for decadal prediction. The use of machine learning techniques, particularly in climate science, has been growing, demonstrating their effectiveness in pattern recognition, teleconnection analysis, and climate variability investigations. This study leverages the power of machine learning to address the challenges posed by limited observational data and to uncover novel insights into sea level predictability.

This study uses two preindustrial control (piControl) simulations from the GFDL SPEAR coupled global climate model to analyze North Atlantic sea level variability. SPEAR_LO and SPEAR_MED have the same ocean resolution but different atmospheric resolutions (1° and 0.5°, respectively). 5000 years of annual mean sea level anomalies (SLA) from each simulation (model years 201–2700, after removing the first 200 years for spin-up) are used. The data is linearly detrended to remove model drift. A self-organizing map (SOM) is employed as an unsupervised machine learning technique to classify the high-dimensional SLA datasets. The SOM organizes high-dimensional data into a lower-dimensional grid (nodes), where nearby nodes represent similar patterns. The study uses a 3x4 SOM grid. The robustness of the SOM is tested using different grid sizes. Composite SLA patterns are generated for each node to identify dominant patterns of variability. To assess multiyear-to-decadal predictability, a SOM-based lagged transition probability framework is used. This involves calculating the probability of each node transitioning to another node (including itself) after 1 to 30 years. This allows for the tracing of temporal SLA evolution and identification of high predictability periods based on significantly higher or lower than normal probabilities compared to climatological frequencies. A model-analog method is used for decadal predictions based on the SOM framework. For each node in the SOM from the piControl simulations, a set of analogs from the simulations is selected based on similarity to the SLA patterns from an observationally constrained SPEAR reanalysis dataset (with the linear trend removed). These analogs and their subsequent evolution are composited to form the predictions. The skill of the model-analog method is compared to that of initialized decadal hindcasts from SPEAR, using the anomaly correlation coefficient (ACC). Tide gauge records along the US Northeast Coast are used for additional evaluation of the predictions. These records are detrended and low-pass filtered to focus on decadal variability. A 15-year high-pass Butterworth filter is applied to the SLA data to isolate shorter-timescale variability, and the SOM analysis is repeated to explore short-term predictability. The results from the high-pass filtered SLA are then compared to satellite altimetry data to assess the consistency of the findings.

The SOM analysis of the piControl simulations identified dominant patterns of North Atlantic SLA variability strongly linked to different phases of the AMOC. Nodes [3,1] and [1,3] represent mature negative and positive AMOC phases, respectively, with distinct SLA patterns along the US East Coast. Nodes [1,1] and [3,4] show distinct loadings along the US Southeast Coast. The lagged transition probability analysis revealed that the persistence and directional transition of SLA patterns (anticlockwise tendency) are attributable to low-frequency buoyancy-driven AMOC variability, leading to decadal predictability lasting 15-20 years. Analysis of 15-year high-pass filtered SLA patterns revealed additional short-term predictability (4-5 years), possibly due to ocean memory, transient-eddy feedback, and slow gyre circulation adjustments triggered by stochastic forcing. Composites of satellite altimeter data mapped onto the SOM from high-pass filtered SLA showed similarities to the piControl simulations, although discrepancies exist due to the shorter observational record and model limitations. The persistence of node [1,1] in satellite observations suggests an accelerated increase in coastal sea level in the US Southeast since 2015. Analysis of the SPEAR large ensemble indicates limited impact of external forcing on short-timescale SLA variability. The model-analog predictions show comparable skill to initialized decadal hindcasts for large-scale SLA patterns and low-frequency coastal SLA variations, especially when starting from certain states like mature AMOC states. The SOM-based model-analog method demonstrated skill in predicting low-frequency US Northeast coastal SLA up to 5-8 years, highlighting the linkage between large-scale SLA patterns and coastal variations. The study found that the model-analog method performed better than hindcasts in the first few years at tide gauge stations in the Northeast.

The findings demonstrate the potential of a SOM-based framework to effectively assess multiyear-to-decadal predictability of North Atlantic sea level. The strong link between identified SLA patterns and AMOC phases supports the importance of AMOC variability in driving decadal sea level changes. The identification of both decadal and shorter-term predictability sources expands our understanding of the complexities of sea level dynamics. The comparable prediction skill of the model-analog method relative to initialized hindcasts shows its promise as a cost-effective alternative for long-lead predictions. The success in predicting low-frequency coastal SLA variations using large-scale patterns underscores the importance of considering spatial interactions in forecasting.

This study provides a novel framework for assessing North Atlantic sea level predictability, highlighting the roles of both low-frequency AMOC variability and shorter-term processes. The SOM-based model-analog approach demonstrates comparable skill to initialized hindcasts for long-lead predictions, offering a cost-effective method for early estimation of future sea levels. Future research should investigate enhancing prediction skill by improving model representation of AMOC variability and incorporating additional processes influencing coastal sea level. Further research could also explore the impact of different horizontal resolutions on the predictability findings.

The study relies on climate model simulations, which have inherent limitations in representing all processes affecting sea level. Model biases and uncertainties may affect the generalizability of findings. The limited duration of satellite altimetry data restricts the full assessment of longer-timescale variability. The analysis focuses on large-scale patterns, and further research should examine finer-scale variability and its predictability. The use of tide gauge data for regional verification is constrained by the availability and quality of these records. The study's focus on internal variability means that external forcings not included in the analysis could influence sea level predictions.