Skillful multiyear to decadal predictions of sea level in the North Atlantic Ocean and U.S. East Coast

Sea level change threatens coastal communities through flooding and inundation, with impacts spanning timescales from hours to decades. While short-term variability is dominated by tides, waves, and storms, seasonal to multidecadal changes are linked to large-scale ocean dynamics and climate variability. Anthropogenic warming drives a long-term, global sea-level rise through thermal expansion and land ice melt. In the North Atlantic, winds, the Gulf Stream, and the Atlantic Meridional Overturning Circulation (AMOC) exert strong control on sea level at interannual to longer timescales. The U.S. East Coast is a hotspot where dynamic sea level changes associated with AMOC/Gulf Stream variations can amplify the global mean rise. Despite the societal need for forecasts beyond seasonal scales, most prior work emphasized seasonal predictions, with relatively limited attention to multiyear-to-decadal horizons. Given demonstrated decadal predictability of AMOC and North Atlantic heat content, this study investigates the multiyear to decadal predictability and prediction skill of North Atlantic and U.S. East Coast sea level using observations, control simulations, and initialized decadal hindcasts.

Prior studies have shown higher seasonal sea level prediction skill along the U.S. West Coast than the East Coast due to stronger ENSO influence. Numerous works demonstrated decadal predictability of the AMOC and North Atlantic heat content, suggesting potential for sea level prediction at multiyear scales. AMOC changes modulate dynamic sea level along the U.S. East Coast via geostrophic balance, with weakening AMOC/Gulf Stream raising coastal sea level. Observational and modeling studies have identified accelerated sea level rise since ~2010 along the U.S. Southeast and Gulf Coasts and links to large-scale ocean heat divergence. These literatures motivate exploring multiyear-to-decadal sea level predictability tied to both external radiative forcing and internal AMOC variability.

The study combines: (1) a multi-millennial preindustrial control simulation using GFDL SPEAR_LO to diagnose potential predictability; (2) initialized retrospective decadal hindcasts/forecasts (20-member ensembles; 10-year integrations) initialized annually (1961–2020) from a SPEAR-based reanalysis (SPEAR_atm_sst_restore) with atmospheric fields restored to JRA-55 and SSTs to ERSSTv5; (3) large ensemble simulations to estimate externally forced signals via signal-to-noise maximizing EOFs; and (4) observational verification using satellite altimetry (Copernicus/AVISO) and tide-gauge records (UHSLC) with inverted barometer correction (using JRA-55 SLP). Predictability identification uses the Average Predictability Time (APT) method, which maximizes the time-integrated fraction of predictable variance across lead times. For initialized hindcasts, ensemble forecast variance and climatological variance are computed directly across leads to solve a generalized eigenvalue problem that yields spatial components maximizing APT and associated time series. For the control run (single realization), a linear regression model estimates lagged covariances for the APT eigenproblem; potential predictability is evaluated via squared multiple correlation R^2 as a function of lead. Statistical significance is assessed with Monte Carlo methods. Externally forced signals are removed from model fields using signal-to-noise maximizing EOFs derived from the SPEAR_LO large ensemble; for tide gauges, linear trends are removed to isolate internal variability. Skill is verified by projecting observations onto model APT spatial patterns and computing anomaly correlations versus hindcast component time series as a function of lead. Additional diagnostics include lagged regressions of AMOC streamfunction against APT components to attribute predictability sources to AMOC mature or transition states. Regional verification along the U.S. East Coast aggregates tide gauges into Northeast, Mid-Atlantic, and Southeast regimes, comparing total and detrended (internal) anomalies and computing lead-dependent skill.

- In the control (perfect-model) context, the leading predictable North Atlantic sea level component (linked to AMOC mature phase) exhibits multidecadal variability (∼25–40 years) and potential predictability of about 7 years (R^2-based). A second component (tripole pattern; AMOC transition phase) has lower predictability, significant up to ∼4 years. - In initialized hindcasts, three predictable components emerge: 1) APT1: A basin-scale upward trend (forced steric sea level rise) with high spatial coherence across the North Atlantic, predictable up to about 10 years. This pattern strongly correlates (spatial r ≈ 0.75) with externally forced variability from large ensembles. 2) APT2: AMOC mature-state-related pattern resembling control APT1, predictable up to ∼5 years; reanalysis APT2 time series is strongly anti-correlated with AMOC strength (r ≈ −0.85). 3) APT3: Tripole sea level pattern associated with AMOC transition between phases, predictable to ∼3 years and correlated with the second streamfunction EOF in reanalysis. - Along the U.S. East Coast: - The trend-like rise is predictable about a decade in advance everywhere, with stronger trend-like skill toward the Northeast due to combined external forcing and AMOC-weakening-induced dynamic SLR. - Detrended internal variability shows regime-dependent skill: Northeast composite predictable up to ∼4 years; Southeast up to ∼3 years; Mid-Atlantic shows limited multiyear skill due to noisy interannual variability. - Tide-gauge verification shows Northeast detrended sea level covaries with AMOC-mature-state component (APT2), while Southeast covaries more with the AMOC-transition component (APT3) and exhibits accelerated SLR after 2010 captured by hindcasts at 1–3 year leads. - Realized AMOC-related sea level skill (∼3–5 years, depending on location) is lower than perfect-model potential (∼5–7 years), attributable to model biases and initialization uncertainties. - Hindcasts initialized 1995–2003 captured the post-2005 transition to multi-annual high sea levels in the Northeast; however, the extreme 2009–2010 event was underestimated, likely due to unpredictable NAO contributions on multiyear scales.

The study demonstrates that sea level along the U.S. East Coast is predictable on multiyear-to-decadal horizons due to two key sources: externally forced steric rise from greenhouse gas warming (providing basin-scale, decadal predictability) and internal AMOC variability (providing 3–5 year skill regionally via mature and transition phases). These findings directly address the need for actionable predictions beyond seasonal scales for coastal planning and risk management. The identification of distinct AMOC-linked components clarifies spatial differences in predictability—stronger internal skill in the Northeast (mature-state AMOC influence) and in the Southeast (transition-state influence), with weaker multiyear skill in the Mid-Atlantic. The results also highlight limitations in capturing extreme events tied to atmospheric modes like the NAO. Overall, by linking predictable sea level patterns to physical drivers (external forcing, AMOC phases), the study provides a mechanistic basis for predictive skill and informs where and over what horizons such forecasts can support adaptation decisions.

The paper identifies three predictable North Atlantic sea level components: a forced, basin-wide upward trend predictable at least a decade ahead; and two AMOC-related internal components—the mature-state pattern and a tripole transition-state pattern—predictable to roughly 5 and 3 years, respectively. Along the U.S. East Coast, the trend-like skill increases from south to north due to combined external forcing and AMOC-weakening effects, while detrended multiyear skill is strongest in the Northeast (∼4 years) and present in the Southeast (∼3 years), but limited in the Mid-Atlantic. Forecasts indicate continued sea level rise in the U.S. Northeast over the coming decade, with internal anomalies remaining above normal for the next few years. Future work should examine how declining AMOC variability under warming may reduce multiyear predictability, and advance model fidelity (e.g., higher-resolution ocean, inclusion of land ice, river discharge, and vertical land motion) and multi-model comparisons to improve forecast accuracy and utility for coastal risk management.

- Model structural limitations: SPEAR_LO lacks a land ice component (cannot represent Greenland melt impacts on AMOC, dynamic and static sea level), omits tides and vertical land motion (subsidence/uplift), and underestimates river discharge contributions along coasts. - Resolution constraints: 1° ocean resolution leads to Gulf Stream separation bias and poorly resolved coasts and shelf breaks, reducing fidelity of regional sea level variability and damping eddy-driven variance. - Magnitude biases: Underestimated SLR magnitude along the U.S. East Coast relative to tide gauges affects trend-like skill estimates. - Initialization and model biases: Reduce realized AMOC-related skill from perfect-model potential (5–7 years) to ∼3–5 years. - Predictability limits: Atmospheric modes (e.g., NAO) that contribute to extreme events (e.g., 2009–2010) are not predictable at multiyear to decadal horizons. - Potential future decline in AMOC variability under warming may reduce multiyear sea level predictability compared to present-day estimates.