Coastal communities face substantial risks from long-term sea level rise and decadal sea level variations, with the North Atlantic and U.S. East Coast being particularly vulnerable under changing climates. Employing a self-organizing map-based framework, we assess the North Atlantic sea level variability and predictability using 5000-year sea level anomalies (SLA) from two preindustrial control model simulations. Preferred transitions among patterns of variability are identified, revealing long-term predictability on decadal timescales related to shifts in Atlantic meridional overturning circulation phases. Combining this framework with model-analog techniques, we demonstrate prediction skill of large-scale SLA patterns and low-frequency coastal SLA variations comparable to that from initialized hindcasts. Moreover, additional short-term predictability is identified after the exclusion of low-frequency signals, which arises from slow gyre circulation adjustment triggered by the North Atlantic Oscillation-like stochastic variability. This study highlights the potential of machine learning to assess sources of predictability and to enable long-term climate prediction.

沿海社区面临着长期海平面上升和年代际海平面变化的巨大风险，其中北大西洋和美国东海岸在气候变化下尤其脆弱。采用基于地图的自组织框架，我们使用来自两个前工业化控制模型模拟的 5000 年海平面异常 （SLA） 来评估北大西洋海平面的变化和可预测性。确定了变率模式之间的首选过渡，揭示了与大西洋经向翻转环流阶段的变化相关的年代际时间尺度上的长期可预测性。将该框架与模型模拟技术相结合，我们展示了大规模 SLA 模式和低频沿海 SLA 变化的预测技能，可与初始化的后报相媲美。此外，在排除低频信号后，确定了额外的短期可预测性，低频信号是由北大西洋振荡样随机变率触发的缓慢环流调整引起的。这项研究强调了机器学习在评估可预测性来源和实现长期气候预测方面的潜力。