Select a result to preview

AI赋能海洋数值模式研究

#AI #数值模式 #海洋 #机器学习

人工智能赋能海洋数值模拟：从次网格参数化到数字孪生海洋的范式革新

摘要

海洋数值模拟作为理解地球气候系统、预测海洋环境变化以及支持海洋经济活动的核心工具，长期以来面临着多尺度物理过程解析与计算资源限制之间的根本性矛盾。随着人工智能（AI），特别是深度学习（DL）技术的迅猛发展，海洋科学领域正在经历一场深刻的范式转移。本研究报告旨在全面、详尽地探讨人工智能方法在传统海洋数值模式中的多重作用，涵盖次网格过程参数化（Subgrid-Scale Parameterization）、计算加速与代理模型（Surrogate Modeling）、数据同化（Data Assimilation）、偏差订正与降尺度（Bias Correction and Downscaling）以及物理融合的机器学习架构（Physics-Informed Machine Learning）。

通过对前沿文献的深度综合分析，本报告指出，AI并非仅仅是传统模式的替代者，而是作为一种强大的“计算协处理器”和“物理发现引擎”，正在重塑海洋学的研究方法。具体而言，基于方程发现的参数化方案解决了涡旋能量反向级联的物理表征难题；基于Transformer和图神经网络的代理模型实现了区域环流模拟数百倍的加速；而端到端的深度学习数据同化框架（如4DVarNet）则突破了传统伴随模式的计算瓶颈。然而，这一融合过程仍面临着模型可解释性、长期积分稳定性以及极端气候背景下泛化能力等严峻挑战。本报告主张，未来的发展方向在于构建“混合物理-AI系统”（Hybrid Physics-AI Systems），即在保留物理守恒律硬约束的前提下，充分利用数据驱动方法的非线性映射能力，从而实现构建高保真、实时交互的“海洋数字孪生”（Digital Twins of the Ocean）的宏伟目标。

1. 引言：海洋数值模拟的尺度困境与AI的崛起

1.1 物理海洋学的多尺度挑战

海洋是一个典型的多尺度、非线性流体动力系统。其动力过程的空间尺度跨越了从毫米级的分子耗散到数千公里的盆地尺度环流，时间尺度则覆盖了从秒级的湍流脉动到数千年的深海热盐环流循环。这种巨大的时空跨度构成了海洋数值模拟的核心挑战。

传统的海洋数值模式基于地球物理流体动力学（GFD）的基本方程组——即在旋转球体上的纳维-斯托克斯（Navier-Stokes）方程组，辅以热力学方程和状态方程。为了在计算机上求解这些方程，必须对连续的海洋流体进行离散化处理，将其划分为网格。然而，受限于当前的计算能力（即便是最强大的超级计算机），全球气候模式（Climate Models）通常只能采用较为粗糙的网格分辨率（例如 100公里或 25公里）。[1][2]

这种离散化导致了严重的“截断误差”。凡是小于网格尺度的物理过程——例如中尺度涡（Mesoscale Eddies, 10-100km）、亚中尺度锋面（Submesoscale Fronts, 1-10km）、内波破碎以及垂直湍流混合——都无法被模式直接求解（Resolve）。然而，这些未被解析的次网格过程（Subgrid-Scale, SGS）对大尺度环流有着至关重要的反馈作用。例如，中尺度涡包含了海洋中超过90%的动能，并且是热量、盐度和碳向极地输送的关键载体。如果在模式中忽略它们，模拟出的西边界流（如墨西哥湾流、黑潮）将会过弱，且延伸路径错误，进而导致巨大的模拟偏差 。[3][4]

为了解决这一问题，传统方法依赖于“参数化”（Parameterization），即利用已解析的大尺度变量（如平均流速、平均温度梯度）来估算未解析的小尺度过程的统计效应。经典的Gent-McWilliams (GM) 参数化方案就是为了模拟涡旋引起的示踪物输运而设计的。然而，这些传统参数化方案大多基于简化的物理假设和有限的观测经验公式，往往难以捕捉复杂的流体非线性相互作用，导致模式存在系统性偏差 。[5][6]

1.2 科学机器学习（SciML）的范式介入

在摩尔定律逐渐失效、单纯依靠堆砌硬件难以无限提高分辨率的背景下，科学机器学习（Scientific Machine Learning, SciML）作为一种新的研究范式应运而生。与传统的“黑箱”AI不同，SciML强调将领域知识（Domain Knowledge）、物理约束（Physical Constraints）与机器学习强大的数据拟合能力相结合 。[1:1]

AI在海洋数值模拟中的介入并非偶然，而是基于海量数据的积累（卫星遥感、Argo浮标、高分辨率数值模拟数据）与深度学习算法（如卷积神经网络CNN、Transformer、神经算子）成熟的共同驱动。AI在此领域的作用主要体现在三个维度的突破：

本报告将依据上述逻辑框架，深入剖析AI在海洋数值模式各个环节的具体应用机制、技术细节及其实际效果。

2. 次网格过程参数化的智能化重构

次网格过程参数化是连接微观湍流与宏观环流的桥梁，也是目前AI介入最深入、理论最成熟的领域。传统的参数化方案往往假设小尺度过程仅仅起到耗散作用（如涡粘性），但实际上，海洋中的能量级联是双向的——能量不仅从大尺度向小尺度传递（正向级联），小尺度涡旋也会组织起来驱动大尺度流动（反向级联或能量靠背，Energy Backscatter）。AI方法在捕捉这种反向级联效应方面展现出了超越传统理论的潜力。

2.1 中尺度涡动量通量的“方程发现”与深度学习

中尺度涡对大尺度平均流的动量输运是维持全球海洋环流形态的关键。在粗分辨率模式中，这种输运必须被准确参数化。

2.1.1 Zanna-Bolton (ZB20) 方案：从数据中发现物理定律

Zanna和Bolton (2020) 的工作是该领域的里程碑。他们没有直接使用深度神经网络作为一个不可解释的黑箱来预测次网格应力，而是采用了一种“方程发现”（Equation Discovery）的方法，具体使用了相关向量机（Relevance Vector Machine, RVM）。RVM是一种稀疏贝叶斯学习算法，能够从预定义的大量候选物理项（如剪切变形、涡度梯度等）中筛选出最能解释数据的几项，组合成解析表达式 。[8:1][13]

通过分析高分辨率CM2.6气候模式的数据，他们发现次网格涡动量通量（Eddy Momentum Flux）主要与大尺度流场的变形张量（Deformation Tensor）和涡度梯度的乘积成正比。其核心发现可以用以下数学形式概括（简化版）：

∇⋅τeddy∝κ∇⋅(D∇ζ¯)

其中，τeddy 是次网格应力，D 是变形率，ζ¯ 是大尺度涡度。这一发现具有极其重要的物理意义：它揭示了涡旋通过“负粘性”机制向平均流注入能量，从而维持了西边界流的强度。当将这一解析方程植入到GFDL MOM6海洋模式中时，它显著改善了模式对海流能量分布的模拟能力，减少了长期存在的模拟偏差 。[14][15]

2.1.2 卷积神经网络（CNN）与物理约束的必要性

除了方程发现，直接利用卷积神经网络（CNN）来预测次网格强迫也是一种主流路径。CNN能够提取流场的空间结构特征（Spatial Features），这对于捕捉涡旋的非局地效应（Non-local effects）至关重要。Guillaumin和Zanna (2021) 训练了一个深层CNN，利用局部的流速场、温度场和密度场来预测次网格动量倾向 。[16][17]

然而，研究发现，直接将训练好的高精度CNN植入到海洋模式中进行在线（Online）运行时，往往会导致模式崩溃。这是因为神经网络在离线（Offline）测试中虽然表现优异，但在在线运行时，哪怕微小的预测误差也会通过流体动力学方程反馈放大，导致数值不稳定。

2.2 垂直混合过程：超越KPP方案

海洋上层混合层（Mixed Layer）的深度决定了海气之间的热量和气体交换效率。传统的K-Profile Parameterization (KPP) 方案基于莫宁-奥布霍夫相似理论（Monin-Obukhov Similarity Theory），虽然应用广泛，但在极端风浪条件或复杂层结下的表现并不理想。

2.2.1 ePBL中的神经网络替换

NOAA的地球物理流体力学实验室（GFDL）开发了基于神经网络的垂直混合参数化方案，以增强MOM6模式中的ePBL（energetically constrained Planetary Boundary Layer）方案。研究人员利用二阶矩闭合（Second Moment Closure, SMC）湍流模型产生的高保真数据训练神经网络，让其学习垂直涡扩散系数（Eddy Diffusivity, Kv）随海洋状态（表面风应力、浮力通量、科氏参数、边界层深度）变化的复杂非线性关系 。[18][19]

在这一混合架构中，神经网络并非接管整个垂直混合计算，而是仅替换掉KPP中经验性最强、不确定性最大的“形状函数”（Shape Function）部分。这种设计极其巧妙，因为它保留了ePBL方案原有的能量守恒框架（Energetic Constraints），仅在局部引入AI的非线性拟合能力。

2.2.2 可解释性回归与符号学习

为了进一步提高模型的可信度，Sane等人采用了“两步走”策略：首先训练一个高精度的神经网络来模拟SMC模型的输出，然后利用符号回归（Symbolic Regression）技术分析神经网络的输入输出关系，反推出一个简化的、物理上可解释的代数方程组 。这一过程实际上是将神经网络作为一个“知识蒸馏器”，从复杂数据中提炼出人类可理解的物理规律。这种方法生成的参数化方案既具有数据驱动的高精度，又具有解析公式的计算高效性和可解释性，是未来参数化发展的理想路径。[21]

表 1：传统参数化与AI参数化方案的深度对比

比较维度 传统参数化 (如 GM, KPP) 深度学习/AI参数化 (如 ZB20, Neural-ePBL)
构建基础 理想化理论假设、有限的实验室/现场观测数据 高分辨率数值模拟（LES/DNS）产生的海量“全知”数据
数学形式 简单的代数公式，依赖经验系数调节 复杂的深层神经网络、卷积网络或通过符号回归发现的新方程
非线性能力 较弱，难以表征复杂的流体状态依赖性 极强，能够捕捉高度非线性的流体动力学特征
能量级联 通常仅模拟正向级联（耗散），导致流场过于平滑 能够模拟反向能量级联（Backscatter），增强大尺度流结构
计算开销 极低，几乎不占用额外资源 中等至高，推理过程（Inference）可能需要GPU加速或矩阵运算优化
稳定性 经过数十年调优，数值稳定性极好 容易出现数值不稳定，需要专门的滤波、物理约束或架构设计

3. 代理模型与计算加速：打破摩尔定律的限制

如果说参数化是为了让模式“更准”，那么代理模型（Surrogate Modeling）则是为了让模式“更快”。在传统的海洋预报中，为了获得高分辨率的预报结果，必须求解极其昂贵的偏微分方程组。AI提供了一种绕过繁重计算的捷径——通过学习输入（初始场/边界条件）与输出（预报场）之间的映射关系，构建一个快速的“仿真器”。

3.1 区域海洋模式（ROMS）的AI极速替代

区域海洋模拟系统（ROMS）是全球应用最广泛的近岸海洋模式之一。然而，受限于CFL（Courant-Friedrichs-Lewy）条件的限制，高分辨率ROMS模拟的时间步长极短，计算量巨大。

3.1.1 4D Swin Transformer 架构

最近的研究提出了一种基于4D Swin Transformer的AI代理模型来加速ROMS 。Swin Transformer是一种层级式的视觉Transformer，它通过“移动窗口”（Shifted Windows）机制在局部注意力（捕捉细微的波浪、湍流）和全局注意力（捕捉大尺度的潮汐传播）之间取得了平衡，非常适合处理具有多尺度特征的地球物理流体数据。[9:1][22]

该模型采用了双分辨率策略：

3.1.2 450倍的惊人加速与混合验证

在性能测试中，对于一个覆盖复杂海岸线的12天预报任务，传统的MPI并行ROMS在512个CPU核上需要运行9908秒（约2.75小时）。相比之下，训练好的AI代理模型在单块NVIDIA A100 GPU上仅需22秒即可完成同样的预报任务。这代表了超过450倍的端到端加速 。[9:2][10:1]

为了保证结果的物理可靠性，该系统引入了一个“混合验证模块”。系统会实时检查AI预测结果是否满足质量守恒定律（Conservation of Mass）。如果AI预测的水位变化导致质量残差超过预设阈值（如 5.0×10−4 m/s），系统会自动回退调用传统的ROMS求解器进行校正。这种“人在回路”或“物理在回路”的设计，既享受了AI的极速，又保留了传统模式的可靠性底线 。[10:2][23]

3.2 全球尺度仿真：从GraphCast到FNO

在全球尺度上，DeepMind开发的GraphCast和NVIDIA开发的FourCastNet代表了AI气象/海洋预报的最高水平。

3.3 长期积分的稳定性难题：谱偏差与漂移

尽管AI代理模型在短期预报（0-10天）表现出色，但在尝试进行长达数月甚至数年的气候模拟时，往往会面临**长期漂移（Climate Drift）和谱偏差（Spectral Bias）**的问题。

4. AI驱动的数据同化：突破伴随模式的瓶颈

数据同化（Data Assimilation, DA）是将观测数据融合进数值模式以获得最优初始场的过程。传统的变分同化方法（如4D-Var）虽然精度高，但计算成本极高，且需要开发极其复杂的切线性模式和伴随模式（Adjoint Model）。AI正在从根本上改变这一流程。

4.1 4DVarNet：端到端的变分学习

4DVarNet 是一种基于深度学习的端到端数据同化框架，旨在直接求解变分同化的代价函数 。[11:1][12:1]

4.2 智能化的集合卡尔曼滤波（EnKF）

在集合卡尔曼滤波中，背景误差协方差矩阵（B矩阵） 决定了观测信息如何在空间中传播。

4.3 超分辨率数据同化

传统同化受限于模式网格，无法有效利用高于模式分辨率的观测信息。AI使得超分辨率同化成为可能。通过生成对抗网络（GAN）或扩散模型，可以将低分辨率的模式背景场与稀疏的高分辨率观测相结合，直接生成高分辨率的分析场。这实际上是在同化步骤中通过AI完成了“动力学降尺度”，为预报模式提供了蕴含更多小尺度信息的初始条件 。[37][38]

5. 偏差订正与物理融合：从后处理到硬约束

5.1 偏差订正与降尺度

即便最先进的气候模式也存在系统性偏差（如双赤道辐合带问题、北大西洋冷偏差）。AI提供了强大的后处理工具。

5.2 物理融合机器学习（Physics-Informed ML）：硬约束与软约束

为了保证AI模型输出的物理一致性，必须引入物理约束。这主要分为“软约束”和“硬约束”两种路径。

5.2.1 软约束：物理信息神经网络（PINNs）

PINNs将控制方程（如浅水方程、波动方程）的残差作为惩罚项加入到损失函数中：

Ltotal=Ldata+λLPDE

这种方法在反演问题中表现出色，例如利用稀疏观测重构波浪场或污染物扩散路径 。然而，在复杂的海洋湍流问题中，PINNs面临优化难题。由于物理损失项和数据损失项的梯度性质差异巨大，优化器往往难以找到帕累托最优解，导致物理约束无法严格满足 。[42][43][44][45][46]

5.2.2 硬约束：网络架构设计

鉴于软约束的局限性，海洋学界越来越倾向于使用硬约束，即通过网络架构设计强制满足物理定律。

u=−∂Ψ/∂y,v=∂Ψ/∂x

，根据向量微积分恒等式，无论网络输出什么数值，计算出的流场自动满足无散度条件 ∇⋅u=0。这种方法保证了质量守恒是“精确”的，而非“近似”的 。[47]

研究表明，硬约束对于模型的泛化能力至关重要。一个未受约束的纯数据模型可能在训练数据覆盖的气候态下表现良好，但在全球变暖等从未见过的新情景下会严重违反物理定律；而硬约束模型则能保证基本的物理守恒律在任何气候背景下都不被打破 。[49]

6. 海洋数字孪生（DTOs）：集大成者

AI技术在海洋学中的终极应用形态是海洋数字孪生（Digital Twins of the Ocean, DTOs）。DTO不再是一个静态的数据文件或一次性的模拟，而是一个动态的、可交互的虚拟海洋系统 。[1:2][50]

AI在DTO中扮演着“交互引擎”的角色。传统的数值模式运行太慢，无法支持决策者进行实时的“假设分析”（What-If Scenarios）。

7. 挑战与展望

尽管前景广阔，AI在海洋数值模拟中的应用仍处于起步阶段，面临诸多核心挑战：

结论

人工智能方法正在从边缘的辅助工具，转变为传统海洋数值模式的核心组件。它通过方程发现深化了我们对湍流物理的理解，通过代理模型实现了计算效率的指数级跃升，通过4DVarNet革新了数据利用的方式。

然而，物理学并不会被AI取代。相反，最成功的模型将是那些将纳维-斯托克斯方程的严谨性与神经网络的灵活性完美结合的物理融合AI系统。这种混合范式不仅能解决传统模式无法解决的参数化难题，更为构建高分辨率、实时交互的海洋数字孪生奠定了坚实基础。

参考文献索引：

🔗 正文参考文献 (References)

🧠 思考过程 (Thoughts)

Step 1: 研究框架构建

我正在针对人工智能在海洋数值模式中的应用构建一套完整的逻辑框架。我将重点关注深度学习如何通过参数化亚尺度物理过程、构建高效的代理模型以及优化数据同化流程来提升传统模式的性能。通过这种多维度的分析，我旨在系统性地呈现人工智能在海洋预报和气候模拟中的核心作用。

Step 2: 物理约束与效能平衡

在思考过程中，我识别出物理一致性与模型效率之间的平衡是目前研究的焦点。我正准备深入探究物理信息神经网络（PINNs）如何将流体力学方程融入学习过程，以解决深度学习模型在传统海洋数值框架中可能出现的物理守恒性缺失问题，并评估其在提升模型分辨率和订正预报偏差方面的实际潜力。

Step 3: 后续研究路径

接下来，我将展开广泛的信息检索，优先锁定领域内的前沿综述与权威研究成果。我计划先梳理出人工智能与海洋模式结合的技术图谱，随后深入挖掘如中尺度涡旋参数化、神经网络加速方案以及物理规律约束下的数据驱动方法等具体技术细节，以确保研究深度覆盖到海洋科学研究的最前沿。

Step 4: ... [57][58][59][60][61][62][63][64][65][66][67][68][69][70][71][72][73][74][75][76][77][78][79][80][81][82][83][84]

Step 5: 突破性效能提升与架构创新

我在最新的调研中发现，人工智能在海洋模拟领域的提速潜力巨大。例如，利用特定架构的代理模型，可以将原本耗时数小时的近岸环流预报大幅缩短至数十秒，实现数百倍的加速。同时，我观察到图神经网络和傅里叶神经算子正成为捕捉中尺度涡旋等复杂动力过程的关键技术，在处理空间多尺度问题上表现出比传统线性模型更高的精度。

Step 6: 从黑盒模型向科学机器学习的范式转变

通过对近期研究的综合分析，我注意到一种向“科学机器学习”转化的明显趋势。这种方法正致力于解决纯数据驱动模型在物理一致性上的局限，通过将流体力学方程和守恒律直接嵌入学习逻辑中，尝试消除传统模式中亚尺度过程参数化带来的系统偏差。这种融合不仅提高了模型在极端海况下的可靠性，也为处理稀疏观测数据的同化问题提供了新思路。

Step 7: 核心模块的深度解构与效能评估

接下来，我将针对海洋垂直混合参数化、代理模型加速方案以及混合数据同化系统展开专项剖析。我计划深入探究如何在算法层面上强制执行质量与能量守恒约束，并详细评估这些智能模块在长期气候预报中的泛化能力，以期构建出一套既具备物理可解释性又能满足实时预报需求的高效海洋数值框架。

Step 8: ... [85][86][87][88][89][90][91][92][93][94][95][96][97][98][99][100][101][102][103][104][105][106][107][108][109][110][111][112][113][114][115][116][117][118][119][120][121][122][123][124][125][126][127][128][129][130][131][132][133][134][135][136][137][138][139][140][141][142][143][144][145][146][147][148][149][150][151][152][153][154][155][156][157][158][159][160][161][162][163][164][165][166][167][168][169][170][171][172][173][174][175][176][177][178][179][180][181][182][183][184][185][186][187][188][189][190][191][192][193]

Step 9: 核心动力过程的智能化重塑

我正深入分析如Zanna-Bolton（ZB20）等具体的中尺度参数化方案，这些方案通过捕捉流速梯度与应力张量的关系，旨在替代传统模式中难以闭合的亚尺度过程。同时，我观察到Swin Transformer架构在近岸环流预报中表现出的极高效率，这种基于注意力机制的代理模型不仅实现了显著的算力节省，还能在复杂海岸边界条件下维持预报精度，这标志着代理模型正从理论验证走向实际业务加速。

Step 10: 跨越物理一致性与长期稳定性挑战

我正在重点审视数据驱动模型在长期积分过程中可能出现的“漂移”问题，并探讨如何通过技术手段将质量与能量守恒定律作为“硬约束”嵌入神经网络架构中，而非仅仅依赖损失函数的软引导。通过对比传统4DVar同化与新兴的4DVarNet，我发现深度学习正尝试通过重构梯度算子来解决传统方法在非线性系统下计算成本过高的问题，这种范式转变正逐步化解观测数据稀疏与模型复杂性之间的矛盾。

Step 11: 向物理增强与全球尺度扩展迈进

接下来，我将剖析物理信息神经网络（PINNs）在全球尺度环流模拟中的局限性与突破口，特别是其在处理真实海底地形与大规模算网布局时的适应性。我计划详细调查如何在MOM6等主流海洋模式中具体部署这些深度学习模块，并深度解构硬性物理约束在算法层面的具体实现，以期为构建既具备严谨物理逻辑又拥有极致运行速度的智能化海洋数值框架提供技术路线。

Step 12: ... [194][195][196][197][198][199][200][201][202][203][204][205][206][207][208][209][210][211][212][213][214][215][216][217][218][219][220][221][222][223][224][225][226][227][228][229][230][231][232][233][234][235][236][237][238][239][240][241][242][243][244][245][246][247][248][249][250][251][252][253][254][255][256][257][258][259][260][261][262][263][264]

Step 13: 亚尺度方案的数值稳定性与变量匹配

我正在深入分析深度学习参数化方案在融入传统模式时的稳定性问题，特别是如何通过对输入场进行滤波处理来抑制模型在耦合运行时产生的数值噪音。我同时也在梳理神经网络在替代传统垂直混合过程时，与主流海洋模式对接所需的具体状态变量，以确保数据驱动模块与动力学框架之间的无缝协作和精准重构。

Step 14: 预报漂移治理与物理守恒硬约束

通过对长时效预报结果的审视，我识别出代理模型在积分过程中可能出现的物理失真和漂移现象。我正在思考如何将质量和能量守恒定律从损失函数的辅助引导提升为网络架构中的硬性约束，例如通过让模型预测物理通量而非直接预测状态量，从根本上解决长期模拟中的系统偏差，从而显著提升模型在跨季度预报中的可靠性。

Step 15: 同化误差矩阵优化与全球尺度应用评估

接下来的研究将聚焦于人工智能如何精准修正数据同化中的背景误差协方差矩阵，探索这种混合策略在处理复杂非线性偏差时的优势。我计划重点考察物理信息神经网络在全球大洋环流模拟中的最新进展，评估其在面对真实复杂海底地形及超大规模计算网格时的扩展能力与计算效能，以期打通从区域验证到全球应用的最后环节。

Step 16: ... [265][266][267][268][269][270][271][272][273][274][275][276][277][278][279][280][281][282][283][284][285][286][287][288][289][290]

Step 17: 物理过程闭合的新型演化路径

我正通过方程发现技术，深度解构中尺度动量通量的参数化逻辑，并探究如何利用空间滤波技术增强此类方案在动力框架中的数值稳定性。针对垂直混合过程，我重点评估了如何通过高性能计算架构实现深度学习模块与主流海洋模式的在线耦合，确保模型能实时感知风应力与层结变化，从而显著提升对上层海洋动力结构的刻画精度。

Step 18: 混合架构下的稳定性治理与谱偏置修正

针对代理模型长期积分的漂移顽疾，我正在合成一种集成物理验证的混合预报策略，能够在预报偏差超限时自动触发动力学引擎。我同时发现，引入谱正则化和硬性物理约束技术可以有效修正深度学习中的频率偏置，使模型不仅在统计指标上表现优异，更能在质量与能量守恒方面达到机器精度，这对于维持跨季度预报的物理合理性至关重要。

Step 19: 复杂地形约束下的全球化应用探索

后续我将深入调查物理信息神经网络在全球大洋模拟中处理复杂海岸边界和陆地掩码时的局限性，并寻找突破全球尺度模拟的技术路径。我计划系统性地总结利用神经网络直接预测流场背景误差矩阵的最前沿方法，以最终完善整套智能化海洋数值预报体系，实现从局部海域高效预测向全球大洋精准模拟的跨越。

Step 20: ... [291][292][293][294][295][296][297][298][299][300][301][302][303][304][305]

Step 21: 极致加速与物理安全开关

我发现基于Swin Transformer等架构的代理模型在近岸环流预报中实现了惊人的450倍加速。尤为关键的是，研究者引入了一种物理验证机制：当代理模型的预报误差或质量守恒破坏超过特定阈值时，系统会自动切换回传统动力引擎。这种“物理安全开关”的设计，有效解决了纯数据驱动模型在极端海况下可靠性不足的痛点。

Step 22: 算子重构与硬约束的效能优势

在深入剖析数据同化与后处理模块时，我合成了一套利用神经网络重构变分算子的新路径，这显著降低了传统伴随模型在高维空间下的计算成本。通过对比发现，在架构层面植入“硬性物理约束”比单纯在损失函数中添加惩罚项更能有效抑制长时效模拟中的物理漂移，这为维持跨季度预报的动力一致性提供了技术标准。

Step 23: 从区域验证向全域模拟的系统集成

目前我已经完成了从亚尺度参数化、高效代理模型到同化误差订正的全链条信息采集，并确认了物理信息神经网络在全球大洋环流模拟中面临的地形适应性瓶颈。接下来的行动中，我将开始对这些碎片化的前沿技术进行系统性集成，构建一套涵盖混合架构、稳定性治理及全球化挑战的深度研究报告，旨在清晰呈现人工智能重塑海洋模式的完整演进图景。

思考过程参考文献

  1. Crafting the Future: Machine learning for ocean forecasting - Reports ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

  2. Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model - Andrew Slavin Ross ↩︎ ↩︎

  3. Implementation and Evaluation of a Machine Learned Mesoscale Eddy Parameterization Into a Numerical Ocean Circulation Model - the NOAA Institutional Repository ↩︎

  4. A stable implementation of a data-driven scale-aware mesoscale parameterization - arXiv ↩︎ ↩︎

  5. An Analysis of Deep Learning Parameterizations for Ocean Subgrid Eddy Forcing - arXiv ↩︎ ↩︎

  6. Physics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations | National Science Review | Oxford Academic ↩︎ ↩︎

  7. Bathymetry-Aware Mesoscale Eddy Parameterizations across Upwelling Slope Fronts: A Machine Learning–Augmented Approach in - AMS Journals ↩︎ ↩︎

  8. Data-Driven Equation Discovery of Ocean Mesoscale Closures - the NOAA Institutional Repository ↩︎ ↩︎ ↩︎

  9. Accelerate Coastal Ocean Circulation Model with AI Surrogate - arXiv ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

  10. A Fast AI Surrogate for Coastal Ocean Circulation Models | Request PDF - ResearchGate ↩︎ ↩︎ ↩︎ ↩︎

  11. Convolutional neural networks for sea surface data assimilation in operational ocean models: test case in the Gulf of Mexico - OS ↩︎ ↩︎ ↩︎

  12. 4DVarNet-SSH: end-to-end learning of variational interpolation schemes for nadir and wide-swath satellite altimetry - GMD ↩︎ ↩︎ ↩︎ ↩︎

  13. Data‐Driven Equation Discovery of Ocean Mesoscale Closures - ResearchGate ↩︎ ↩︎

  14. [2311.02517] A stable implementation of a data-driven scale-aware mesoscale parameterization - arXiv ↩︎ ↩︎

  15. A Stable Implementation of a Data-Driven Scale-Aware Mesoscale Parameterization ↩︎ ↩︎

  16. Applications of Deep Learning parameterization of Ocean Momentum Forcing - arXiv ↩︎

  17. (PDF) Stochastic‐Deep Learning Parameterization of Ocean Momentum Forcing - ResearchGate ↩︎ ↩︎

  18. Parameterizing Vertical Turbulent Mixing Coefficients for The Ocean Surface Boundary Layer Using Machine Learning | US CLIVAR ↩︎

  19. (PDF) Parameterizing Vertical Mixing Coefficients in the Ocean Surface Boundary Layer Using Neural Networks - ResearchGate ↩︎ ↩︎

  20. Parameterizing Vertical Turbulent Mixing Coefficients In The Ocean Surface Boundary Layer Using Neural Networks ↩︎

  21. Machine Learned Equations for Vertical Mixing in the Ocean Surface Boundary Layer - OSF ↩︎

  22. A Fast AI Surrogate for Coastal Ocean Circulation Models - arXiv ↩︎ ↩︎

  23. Accelerate Coastal Ocean Circulation Model with AI Surrogate - arXiv ↩︎ ↩︎ ↩︎

  24. GraphCast: AI model for faster and more accurate global weather forecasting ↩︎ ↩︎

  25. (PDF) GraphCast: Learning skillful medium-range global weather forecasting ↩︎

  26. Streamlining Ocean Dynamics Modeling with Fourier Neural Operators: A Multiobjective Hyperparameter and Architecture Optimization Approach - MDPI ↩︎ ↩︎

  27. [2404.05768] Streamlining Ocean Dynamics Modeling with Fourier Neural Operators: A Multiobjective Hyperparameter and Architecture Optimization Approach - arXiv ↩︎

  28. Applications of the Fourier neural operator in a regional ocean modeling and prediction ↩︎

  29. Multi-scale, Long-term Stable, and Physically-consistent AI-based Ocean Modeling and Downscaling ↩︎ ↩︎

  30. Generative AI models capture realistic sea-ice evolution from days to decades - arXiv ↩︎

  31. 4DVarNet-SSH: end-to-end learning of variational interpolation schemes for nadir and wide-swath satellite altimetry - GMD ↩︎ ↩︎

  32. arXiv:2207.01372v1 [eess.IV] 4 Jul 2022 - Archimer ↩︎ ↩︎ ↩︎

  33. Length Scale Analyses of Background Error Covariances for EnKF and EnSRF Data Assimilation - MDPI ↩︎ ↩︎

  34. An Adaptive Variance Adjustment Strategy for a Static Background Error Covariance Matrix—Part I: Verification in the Lorenz-96 Model - MDPI ↩︎

  35. Convolutional neural network-based adaptive localization for an ensemble Kalman filter | Computational and Information Systems Lab - cisl.ucar.edu ↩︎

  36. On Domain Localization in Ensemble-Based Kalman Filter Algorithms in - AMS Journals ↩︎

  37. High-resolution Ensemble Kalman Filter with a low-resolution model using a machine learning super-resolution approach - ECMWF Events (Indico) ↩︎

  38. Advancing Ocean State Estimation with efficient and scalable AI - arXiv ↩︎

  39. Contrastive Learning for Climate Model Bias Correction and Super ... ↩︎ ↩︎

  40. A Hybrid Deep Learning Model for the Bias Correction of SST Numerical Forecast Products Using Satellite Data - MDPI ↩︎ ↩︎

  41. How to systematically develop an effective AI-based bias correction model? - arXiv ↩︎ ↩︎

  42. Bridging ocean wave physics and deep learning: Physics-informed neural operators for nonlinear wavefield reconstruction in real-time - arXiv ↩︎ ↩︎

  43. PHYSICS-INFORMED NEURAL NETWORKS FOR MODELING OCEAN POLLUTANT - arXiv ↩︎ ↩︎

  44. (PDF) Physical Informed Neural Networks for modeling ocean pollutant - ResearchGate ↩︎

  45. Full article: NDAWL-PINN: a new non-dimensionalization and multi-task learning approach for efficient training of physics-informed neural networks to solve the shallow water equations - Taylor & Francis Online ↩︎ ↩︎

  46. Deep learning in the abyss: a stratified Physics Informed Neural Network for data assimilation - arXiv ↩︎ ↩︎

  47. Hard-Constrained Deep Learning for Climate Downscaling - arXiv ↩︎ ↩︎

  48. Physics-Constrained Deep Learning for Accelerating Climate Modeling - kluedo ↩︎

  49. Physics-Constrained Deep Learning Postprocessing of Temperature and Humidity in - AMS Journals ↩︎

  50. Digital twins - Ditto ↩︎ ↩︎

  51. Predicting the changing marine conditions with future digital twins ↩︎ ↩︎

  52. Recent Advancements and Challenges in Artificial Intelligence for Digital Twins of the Ocean ↩︎ ↩︎

  53. 4DVarNet-OFDA | CMEMS - Copernicus Marine Service ↩︎

  54. Deep Learning for Ocean Forecasting: A Comprehensive Review of Methods, Applications, and Datasets - IEEE Xplore ↩︎ ↩︎

  55. Correcting Nonstationary Sea Surface Temperature Bias in NCEP CFSv2 Using Ensemble-Based Neural Networks in - AMS Journals ↩︎

  56. Super-resolution of sea surface temperature with convolutional neural network- and generative adversarial network-based methods | Journal of Water and Climate Change | IWA Publishing ↩︎

  57. arxiv.orgDeep Learning Weather Models for Subregional Ocean Forecasting: A Case Study on the Canary Current Upwelling System - arXiv ↩︎

  58. arxiv.orgAccurate Mediterranean Sea forecasting via graph-based deep learning - arXiv ↩︎

  59. arxiv.orgBridging ocean wave physics and deep learning: Physics-informed neural operators for nonlinear wavefield reconstruction in real-time - arXiv ↩︎

  60. arxiv.orgA Framework for Hybrid Physics-AI Coupled Ocean Models - arXiv ↩︎

  61. agu.confex.comMachine Learning Subgrid-Scale Parameterizations for Earth System Modeling I Oral ↩︎

  62. arxiv.orgAn of Deep Learning Parameterizations for Ocean Subgrid Eddy Forcing - arXiv ↩︎

  63. academic.oup.comPhysics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations | National Science Review | Oxford Academic ↩︎

  64. asross.github.ioBenchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model - Andrew Slavin Ross ↩︎

  65. tc.copernicus.orgDeep learning subgrid-scale parametrisations for short-term forecasting of sea-ice dynamics with a Maxwell elasto-brittle rheology - TC ↩︎

  66. arxiv.orgAccelerate Coastal Ocean Circulation Model with AI Surrogate - arXiv ↩︎

  67. sp.copernicus.orgCrafting the Future: Machine learning for ocean forecasting - Reports ↩︎

  68. sp.copernicus.orgCrafting the Future: Machine learning for ocean forecasting - Reports ↩︎

  69. journals.ametsoc.orgData Assimilation with Machine Learning Surrogate Models: A Case Study with FourCastNet in - AMS Journals ↩︎

  70. frontiersin.orgSurrogate modeling for the climate sciences dynamics with machine learning and data assimilation - Frontiers ↩︎

  71. frontiersin.orgApplications of deep learning in physical oceanography: a comprehensive review - Frontiers ↩︎

  72. os.copernicus.orgConvolutional neural networks for sea surface data assimilation in operational ocean models: test case in the Gulf of Mexico - OS ↩︎

  73. ieeexplore.ieee.orgDeep Learning for Ocean Forecasting: A Comprehensive Review of Methods, Applications, and Datasets - IEEE Xplore ↩︎

  74. mdpi.comEarth System Modeling, Data Assimilation, Artificial Intelligence, Deep Learning and Ocean Information Engineering II - MDPI ↩︎

  75. arxiv.orgReconstructing the Tropical Pacific Upper Ocean using Online Data Assimilation with a Deep Learning model - arXiv ↩︎

  76. researchgate.net(PDF) A Deep Learning‐Based Bias Correction Method for Predicting Ocean Surface Waves in the Northwest Pacific Ocean - ResearchGate ↩︎

  77. s3.us-east-1.amazonaws.comContrastive Learning for Climate Model Bias Correction and Super-Resolution - Amazon S3 ↩︎

  78. journals.ametsoc.orgBias and Trend Correction of Precipitation Datasets to Force Ocean Models in - AMS Journals ↩︎

  79. gmd.copernicus.orgImpact of increased resolution on long-standing biases in HighResMIP-PRIMAVERA climate models - GMD ↩︎

  80. arxiv.org[2211.07555] Contrastive Learning for Climate Model Bias Correction and Super-Resolution ↩︎

  81. arxiv.orgPHYSICS-INFORMED NEURAL NETWORKS FOR MODELING OCEAN POLLUTANT - arXiv ↩︎

  82. pubs.aip.orgPhysics-informed neural network with pretraining optimization for ocean acoustic field prediction - AIP Publishing ↩︎

  83. ieeexplore.ieee.orgCL-PINN: A Physics-Informed Neural Networks Framework for Equatorial Tsunami Wave Propagation in Geophysical Ocean - IEEE Xplore ↩︎

  84. mdpi.comMachine Learning in Fluid Dynamics—Physics-Informed Neural Networks (PINNs) Using Sparse Data: A Review - MDPI ↩︎

  85. archimer.ifremer.frObservation-only deep learning for gappy satellite-derived ocean colour data using 4DVarNet - Archimer ↩︎

  86. ieeexplore.ieee.orgObservation-Only Deep Learning for Gappy Satellite-Derived Ocean Color Data Using 4DVarNet | IEEE Journals & Magazine ↩︎

  87. marine.copernicus.eu4DVarNet-OFDA | CMEMS - Copernicus Marine Service ↩︎

  88. gmd.copernicus.org4DVarNet-SSH: end-to-end learning of variational interpolation schemes for nadir and wide-swath satellite altimetry - GMD ↩︎

  89. gmd.copernicus.org4DVarNet-SSH: end-to-end learning of variational interpolation schemes for nadir and wide-swath satellite altimetry - GMD ↩︎

  90. mdpi.comKalman Filter and Its Application in Data Assimilation - MDPI ↩︎

  91. events.ecmwf.intHigh-resolution Ensemble Kalman Filter with a low-resolution model using a machine learning super-resolution approach - ECMWF Events (Indico) ↩︎

  92. gmd.copernicus.orgA gridded air quality forecast through fusing site-available machine learning predictions from RFSML v1.0 and chemical transport model results from GEOS-Chem v13.1.0 using the ensemble Kalman filter - GMD ↩︎

  93. centaur.reading.ac.ukA Multi‐Model Ensemble Kalman Filter for Data Assimilation and Forecasting - CentAUR ↩︎

  94. aimsciences.orgMachine learning-based conditional mean filter: A generalization of the ensemble Kalman filter for nonlinear data assimilation - American Institute of Mathematical Sciences ↩︎

  95. climate.govAI methods enhance rainfall and ocean forecasting in climate model ↩︎

  96. arxiv.orgAdvancing Ocean State Estimation with efficient and scalable AI - arXiv ↩︎

  97. sp.copernicus.orgCrafting the Future: Machine Learning for Ocean Forecasting - Reports ↩︎

  98. bg.copernicus.orgHybrid machine learning data assimilation for marine biogeochemistry - BG ↩︎

  99. arxiv.org[2510.22676] A Framework for Hybrid Physics-AI Coupled Ocean Models - arXiv ↩︎

  100. mdpi.comApplication of Hybrid Data Assimilation Methods for Mesoscale Eddy Simulation and Prediction in the South China Sea - MDPI ↩︎

  101. repository.library.noaa.govUsing Machine Learning for Data Assimilation, Model Physics, and Post-Processing model outputs - the NOAA Institutional Repository ↩︎

  102. ideas.repec.orgHigher-order error estimates for physics-informed neural networks approximating the primitive equations - IDEAS/RePEc ↩︎

  103. usclivar.orgHow to model the Ocean Circulation Eddies using Artificial Neural Networks (OCEANN) ↩︎

  104. arxiv.org[2209.11929] Higher-Order error estimates for physics-informed neural networks approximating the primitive equations - arXiv ↩︎

  105. researchgate.netHigher-Order Error estimates for physics-informed neural networks approximating the primitive equations - ResearchGate ↩︎

  106. pubmed.ncbi.nlm.nih.govPredicting ocean pressure field with a physics-informed neural network - PubMed ↩︎

  107. asmedigitalcollection.asme.orgPhysics-Informed Neural Networks for Modeling Linear Waves - ASME Digital Collection ↩︎

  108. pubs.aip.orgStable adaptive training for physics-informed neural networks in acoustic wave propagation ↩︎

  109. ieeexplore.ieee.orgCL-PINN: A Physics-Informed Neural Networks Framework for Equatorial Tsunami Wave Propagation in Geophysical Ocean - IEEE Xplore ↩︎

  110. arxiv.org[2409.19851] Inference of water waves surface elevation from horizontal velocity components using physics informed neural networks (PINN) - arXiv ↩︎

  111. mdpi.comAn Improved PINN Algorithm for Shallow Water Equations Driven by Deep Learning - MDPI ↩︎

  112. ieeexplore.ieee.orgCL-PINN: A Physics-Informed Neural Networks Framework for Equatorial Tsunami Wave Propagation in Geophysical Ocean - IEEE Xplore ↩︎

  113. arxiv.orgPhysics-informed neural networks for tsunami inundation modeling - arXiv ↩︎

  114. tandfonline.comFull article: NDAWL-PINN: a new non-dimensionalization and multi-task learning approach for efficient training of physics-informed neural networks to solve the shallow water equations - Taylor & Francis Online ↩︎

  115. arxiv.org[2104.00615] Physics-informed neural networks for the shallow-water equations on the sphere - arXiv ↩︎

  116. arxiv.orgA Framework for Hybrid Physics-AI Coupled Ocean Models - arXiv ↩︎

  117. journals.ametsoc.orgPhysics-Constrained Deep Learning Postprocessing of Temperature and Humidity in - AMS Journals ↩︎

  118. amazon.sciencePhysics-constrained machine learning for scientific computing - Amazon Science ↩︎

  119. news.ucsc.eduRegional ocean dynamics can be better emulated with AI models - UC Santa Cruz - News ↩︎

  120. climatechange.aiHybrid Physical Models | Climate Change AI ↩︎

  121. researchgate.netA Fast AI Surrogate for Coastal Ocean Circulation Models | Request PDF - ResearchGate ↩︎

  122. arxiv.orgAccelerate Coastal Ocean Circulation Model with AI Surrogate - arXiv ↩︎

  123. themoonlight.io[Literature Review] A Fast AI Surrogate for Coastal Ocean Circulation Models - Moonlight ↩︎

  124. github.comspatialdatasciencegroup/coastal_circulation_ai_surrogate - GitHub ↩︎

  125. arxiv.org[2410.14952] Accelerate Coastal Ocean Circulation Model with AI Surrogate - arXiv ↩︎

  126. mdpi.comStreamlining Ocean Dynamics Modeling with Fourier Neural Operators: A Multiobjective Hyperparameter and Architecture Optimization Approach - MDPI ↩︎

  127. dspace.mit.eduMIT Open Access Articles Evaluation of Deep Neural Operator Models Toward Ocean Forecasting ↩︎

  128. arxiv.org[2404.05768] Streamlining Ocean Dynamics Modeling with Fourier Neural Operators: A Multiobjective Hyperparameter and Architecture Optimization Approach - arXiv ↩︎

  129. frontiersin.orgApplications of the Fourier neural operator in a regional ocean modeling and prediction ↩︎

  130. sp.copernicus.orgMachine Learning for Ocean Forecasting” By Patrick Heimbach, Fearghal O'Donncha, Jose Maria ↩︎

  131. journals.ametsoc.orgHybrid Numerical Weather Prediction: Downscaling GraphCast AI Forecasts for Downslope Windstorms in - AMS Journals ↩︎

  132. deepmind.googleGraphCast: AI model for faster and more accurate global weather forecasting ↩︎

  133. epic.noaa.govWoFSCast: A GraphCast-based emulator for the Warn-on-Forecast System ↩︎

  134. meteomatics.comFourCastNet and GraphCast Now Available in the Weather API - Meteomatics ↩︎

  135. ditto-oceandecade.orgDigital twins - Ditto ↩︎

  136. journals.ametsoc.orgLong-Term Predictions of Loop Current Eddy Evolutions Using OceanNet: A Fourier Neural Operator–Based Data-Driven Ocean Emulator in - AMS Journals ↩︎

  137. frontiersin.orgFuture digital twins: emulating a highly complex marine biogeochemical model with machine learning to predict hypoxia - Frontiers ↩︎

  138. pml.ac.ukPredicting the changing marine conditions with future digital twins ↩︎

  139. ieeexplore.ieee.orgDeep Learning for Ocean Forecasting: A Comprehensive Review of Methods, Applications, and Datasets - IEEE Xplore ↩︎

  140. mdpi.comThe State of the Art in Deep Learning Applications, Challenges, and Future Prospects: A Comprehensive Review of Flood Forecasting and Management - MDPI ↩︎

  141. arxiv.orgCERA: A Framework for Improved Generalization of Machine Learning Models to Changed Climates - arXiv ↩︎

  142. cpo.noaa.govAI methods enhance rainfall and ocean forecasting in climate model ↩︎

  143. pnas.orgCan AI weather models predict out-of-distribution gray swan tropical cyclones? - PNAS ↩︎

  144. pmc.ncbi.nlm.nih.govArtificial intelligence for geoscience: Progress, challenges, and perspectives - PMC ↩︎

  145. mdpi.comA Review of Machine Learning Applications in Ocean Color Remote Sensing - MDPI ↩︎

  146. academic.oup.comUnlocking the potential of deep learning for marine ecology: overview, applications, and outlook - Oxford Academic ↩︎

  147. spo.nmfs.noaa.govMachine Learning to Improve Marine Science for the Sustainability of Living Ocean Resources - Scientific Publications Office ↩︎

  148. pmc.ncbi.nlm.nih.govApplications of Machine Learning in Chemical and Biological Oceanography - PMC ↩︎

  149. frontiersin.orgApplications of deep learning in physical oceanography: a comprehensive review - Frontiers ↩︎

  150. frontiersin.orgA review of artificial intelligence in marine science - Frontiers ↩︎

  151. mdpi.comRecent Advancements and Challenges in Artificial Intelligence for Digital Twins of the Ocean ↩︎

  152. tos.orgFEATURE ARTICLE • The Promise and Pitfalls of Machine Learning in Ocean Remote Sensing - The Oceanography Society ↩︎

  153. digitalcommons.coastal.edu"Applications of Artificial Intelligence to Improve Coastal Ocean Model" by Dongliang Shen ↩︎

  154. blogs.reading.ac.ukCan We Use Artificial Intelligence To Improve Numerical Models Of The Climate? ↩︎

  155. journals.ametsoc.orgCorrecting Nonstationary Sea Surface Temperature Bias in NCEP CFSv2 Using Ensemble-Based Neural Networks in - AMS Journals ↩︎

  156. mdpi.comA Comparative Evaluation of Two Bias Correction Approaches for SST Forecasting: Data Assimilation Versus Deep Learning Strategies - MDPI ↩︎

  157. mdpi.comA Hybrid Deep Learning Model for the Bias Correction of SST Numerical Forecast Products Using Satellite Data - MDPI ↩︎

  158. gmd.copernicus.orgCustomized deep learning for precipitation bias correction and downscaling - GMD ↩︎

  159. arxiv.orgGlobal Climate Model Bias Correction Using Deep Learning - arXiv ↩︎

  160. journals.ametsoc.orgAlgorithmic Hallucinations of Near-Surface Winds: Statistical Downscaling with Generative Adversarial Networks to Convection - AMS Journals ↩︎

  161. iwaponline.comSuper-resolution of sea surface temperature with convolutional neural network- and generative adversarial network-based methods | Journal of Water and Climate Change | IWA Publishing ↩︎

  162. gmd.copernicus.orgDownscaling atmospheric chemistry simulations with physically consistent deep learning ↩︎

  163. mdpi.comA Transfer Learning-Enhanced Generative Adversarial Network for Downscaling Sea Surface Height through Heterogeneous Data Fusion - MDPI ↩︎

  164. cambridge.orgSea surface height super-resolution using high-resolution sea surface temperature with a subpixel convolutional residual network | Environmental Data Science | Cambridge Core ↩︎

  165. archimer.ifremer.frDeep learning for statistical downscaling of sea states - Archimer ↩︎

  166. journals.ametsoc.orgEnhancing Regional Climate Downscaling through Advances in Machine Learning in - AMS Journals ↩︎

  167. arxiv.orgClimate Downscaling: A Deep-Learning Based Super-resolution Model of Precipitation Data with Attention Block and Skip Connections - arXiv ↩︎

  168. pmc.ncbi.nlm.nih.govAn improved deep learning procedure for statistical downscaling of climate data - PMC ↩︎

  169. ascmo.copernicus.orgDeep learning for statistical downscaling of sea states - ASCMO ↩︎

  170. journals.ametsoc.orgIdentifying and Categorizing Bias in AI/ML for Earth Sciences in - AMS Journals ↩︎

  171. tandfonline.comFull article: Bias correction methods for decadal sea-surface temperature forecasts ↩︎

  172. mdpi.comA Simple Bias Correction Scheme in Ocean Data Assimilation - MDPI ↩︎

  173. arxiv.orgHow to systematically develop an effective AI-based bias correction model? - arXiv ↩︎

  174. academic.oup.comMulti-model approach for statistical downscaling and bias correction of Earth system model outputs for marine ecosystem impact applications - Oxford Academic ↩︎

  175. journals.ametsoc.orgBathymetry-Aware Mesoscale Eddy Parameterizations across Upwelling Slope Fronts: A Machine Learning–Augmented Approach in - AMS Journals ↩︎

  176. arxiv.orgAn of Deep Learning Parameterizations for Ocean Subgrid Eddy Forcing - arXiv ↩︎

  177. repository.library.noaa.govImplementation and Evaluation of a Machine Learned Mesoscale Eddy Parameterization Into a Numerical Ocean Circulation Model - the NOAA Institutional Repository ↩︎

  178. arxiv.orgApplications of Deep Learning parameterization of Ocean Momentum Forcing - arXiv ↩︎

  179. pmc.ncbi.nlm.nih.govPhysics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations - PMC - NIH ↩︎

  180. usclivar.orgParameterizing Vertical Turbulent Mixing Coefficients for The Ocean Surface Boundary Layer Using Machine Learning | US CLIVAR ↩︎

  181. researchgate.net(PDF) Parameterizing Vertical Mixing Coefficients in the Ocean Surface Boundary Layer Using Neural Networks - ResearchGate ↩︎

  182. repository.lsu.eduExploring the use of machine learning to parameterize vertical mixing in the ocean surface boundary layer - LSU Scholarly Repository ↩︎

  183. cesm.ucar.eduParameterizing Vertical Turbulent Mixing Coefficients In The Ocean Surface Boundary Layer Using Neural Networks ↩︎

  184. academic.oup.comPhysics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations | National Science Review | Oxford Academic ↩︎

  185. climatechange.aiIdentifying Spatial Patterns of Biases in Ocean Turbulence Parameterizations using Unsupervised Machine Learning (Papers Track) - Climate Change AI ↩︎

  186. siam.orgMachine Learning for Multiscale Systems: From Turbulence to Climate Prediction - SIAM.org ↩︎

  187. repository.library.noaa.govData-Driven Equation Discovery of Ocean Mesoscale Closures - the NOAA Institutional Repository ↩︎

  188. caos.cims.nyu.eduData-driven parameterizations of ocean mesoscale eddies | Center for Atmosphere Ocean Science | NYU Courant - NYU/CAOS ↩︎

  189. collaborate.princeton.eduA Stable Implementation of a Data-Driven Scale-Aware Mesoscale Parameterization ↩︎

  190. zanna-researchteam.github.ioData-Driven Equation Discovery of Ocean Mesoscale Closures ↩︎

  191. arxiv.org[2311.02517] A stable implementation of a data-driven scale-aware mesoscale parameterization - arXiv ↩︎

  192. sp.copernicus.orgCrafting the Future: Machine learning for ocean forecasting - Reports ↩︎

  193. arxiv.orgContrastive Learning for Climate Model Bias Correction and Super ... ↩︎

  194. par.nsf.govUse of neural networks for stable, accurate and physically consistent parameterization of subgrid atmospheric processes with good performance at reduced precision ↩︎

  195. zanna-researchteam.github.ioData-Driven Equation Discovery of Ocean Mesoscale Closures ↩︎

  196. researchgate.netData‐Driven Equation Discovery of Ocean Mesoscale Closures - ResearchGate ↩︎

  197. online.kitp.ucsb.eduImplementing machine learned parameterizations in climate models - KITP Online Talks ↩︎

  198. osf.ioMachine Learned Equations for Vertical Mixing in the Ocean Surface Boundary Layer - OSF ↩︎

  199. arxiv.orgA stable implementation of a data-driven scale-aware mesoscale parameterization - arXiv ↩︎

  200. gmd.copernicus.orgA high-resolution physical–biogeochemical model for marine resource applications in the northwest Atlantic (MOM6-COBALT-NWA12 v1.0) - GMD ↩︎

  201. cesm.ucar.eduParameterizing Vertical Turbulent Mixing Coefficients In The Ocean Surface Boundary Layer Using Neural Networks ↩︎

  202. researchgate.net(PDF) Parameterizing Vertical Mixing Coefficients in the Ocean Surface Boundary Layer Using Neural Networks - ResearchGate ↩︎

  203. files.cesm.ucar.eduOcean Modeling I Ocean Modeling Basics and CESM Ocean Models - Files ↩︎

  204. workflows.communitySmartSim: Combining Machine Learning and HPC Workflows ↩︎

  205. par.nsf.govUsing Machine Learning at scale in numerical simulations with SmartSim: An application to ocean climate modeling ↩︎

  206. arxiv.orgUsing Machine Learning at Scale in HPC Simulations with SmartSim: - arXiv ↩︎

  207. arxiv.org[2104.09355] Using Machine Learning at Scale in HPC Simulations with SmartSim: An Application to Ocean Climate Modeling - arXiv ↩︎

  208. arxiv.orgA Fast AI Surrogate for Coastal Ocean Circulation Models - arXiv ↩︎

  209. themoonlight.io[Papierüberprüfung] A Fast AI Surrogate for Coastal Ocean Circulation Models - Moonlight ↩︎

  210. github.comspatialdatasciencegroup/CoastalOceanAISurrogate: [IPDPS '25] Accelerate Coastal Ocean Circulation Model with AI Surrogate - GitHub ↩︎

  211. researchgate.netA Fast AI Surrogate for Coastal Ocean Circulation Models | Request PDF - ResearchGate ↩︎

  212. ocean.weather.govarXiv:2502.01797v1 [physics.ao-ph] 3 Feb 2025 - Ocean Weather.Gov ↩︎

  213. arxiv.orgAtmosMJ: Revisiting Gating Mechanism for AI Weather Forecasting Beyond the Year Scale ↩︎

  214. deepmind.googleGraphCast: AI model for faster and more accurate global weather forecasting ↩︎

  215. researchgate.net(PDF) GraphCast: Learning skillful medium-range global weather forecasting ↩︎

  216. researchgate.net(PDF) Optimizing data-driven arctic marine forecasting: a comparative analysis of MariNet, FourCastNet, and PhyDNet - ResearchGate ↩︎

  217. arxiv.org[2202.11214] FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators - arXiv ↩︎

  218. frontiersin.orgOptimizing data-driven arctic marine forecasting: a comparative analysis of MariNet, FourCastNet, and PhyDNet - Frontiers ↩︎

  219. developer.nvidia.comFourCastNet 3 Enables Fast and Accurate Large Ensemble Weather Forecasting with Scalable Geometric ML | NVIDIA Technical Blog ↩︎

  220. gmd.copernicus.orgDo data-driven models beat numerical models in forecasting weather extremes? A comparison of IFS HRES, Pangu-Weather, and GraphCast - GMD ↩︎

  221. arxiv.orgDeep learning in the abyss: a stratified Physics Informed Neural Network for data assimilation - arXiv ↩︎

  222. frontiersin.orgPhysics-informed neural networks to reconstruct surface velocity field from drifter data ↩︎

  223. ieeexplore.ieee.orgPhysics Informed Neural Networks for Modeling Large-Scale Wind Driven Ocean Circulation - IEEE Xplore ↩︎

  224. mdpi.comInvestigation of Physics-Informed Methods for Improving Sea Surface Height Prediction Based on Neural Networks in the South China Sea - MDPI ↩︎

  225. aiforgood.itu.intPhysics-informed machine learning to push the ocean frontier in climate - AI for Good - ITU ↩︎

  226. ml4physicalsciences.github.ioDigital Twin Earth - Coasts: Developing a fast and physics-informed surrogate model for coastal floods via neural operators ↩︎

  227. gmd.copernicus.orgMetrics - Interactive ocean bathymetry and coastlines for simulating the last deglaciation with the Max Planck Institute Earth System Model (MPI-ESM-v1.2) - GMD ↩︎

  228. mdpi.comAn Improved PINN Algorithm for Shallow Water Equations Driven by Deep Learning - MDPI ↩︎

  229. researchgate.net(PDF) Physical Informed Neural Networks for modeling ocean pollutant - ResearchGate ↩︎

  230. arxiv.orgPHYSICS-INFORMED NEURAL NETWORKS FOR MODELING OCEAN POLLUTANT - arXiv ↩︎

  231. mdpi.comStreamlining Ocean Dynamics Modeling with Fourier Neural Operators: A Multiobjective Hyperparameter and Architecture Optimization Approach - MDPI ↩︎

  232. arxiv.orgBridging ocean wave physics and deep learning: Physics-informed neural operators for nonlinear wavefield reconstruction in real-time - arXiv ↩︎

  233. frontiersin.orgApplications of the Fourier neural operator in a regional ocean modeling and prediction ↩︎

  234. dspace.mit.eduMIT Open Access Articles Evaluation of Deep Neural Operator Models Toward Ocean Forecasting ↩︎

  235. arxiv.orgHard-Constrained Deep Learning for Climate Downscaling - arXiv ↩︎

  236. pubs.aip.orgPhysics-informed neural networks with hard and soft boundary conditions for linear free surface waves - AIP Publishing ↩︎

  237. proceedings.mlr.pressLearning Physical Models that Can Respect Conservation Laws ↩︎

  238. mdpi.comWave Equation Modeling via Physics-Informed Neural Networks: Models of Soft and Hard Constraints for Initial and Boundary Conditions - MDPI ↩︎

  239. gmd.copernicus.orgConservation laws in a neural network architecture: enforcing the atom balance of a Julia-based photochemical model (v0.2.0) - GMD ↩︎

  240. jmlr.orgHard-Constrained Deep Learning for Climate Downscaling ↩︎

  241. s3.us-east-1.amazonaws.comAchieving Conservation of Energy in Neural Network Emulators for Climate Modeling - Amazon S3 ↩︎

  242. arxiv.orgA Framework for Hybrid Physics-AI Coupled Ocean Models - arXiv ↩︎

  243. amazon.sciencePhysics-constrained machine learning for scientific computing - Amazon Science ↩︎

  244. journals.ametsoc.orgPhysics-Constrained Deep Learning Postprocessing of Temperature and Humidity in - AMS Journals ↩︎

  245. academic.oup.comPhysics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations | National Science Review | Oxford Academic ↩︎

  246. news.ucsc.eduRegional ocean dynamics can be better emulated with AI models - UC Santa Cruz - News ↩︎

  247. frontiersin.orgA review of artificial intelligence in marine science - Frontiers ↩︎

  248. eos.orgPhysics + Machine Learning Provide a Better Map of Ocean Measurements - Eos.org ↩︎

  249. gmd.copernicus.org4DVarNet-SSH: end-to-end learning of variational interpolation schemes for nadir and wide-swath satellite altimetry - GMD ↩︎

  250. marine.copernicus.eu4DVarNet-OFDA | CMEMS - Copernicus Marine Service ↩︎

  251. gmd.copernicus.org4DVarNet-SSH: end-to-end learning of variational interpolation schemes for nadir and wide-swath satellite altimetry - GMD ↩︎

  252. ostst.aviso.altimetry.fr4DVarNet, an end-to-end learning of variational interpolation schemes: current applications on stellite-derived data and on-goin - OSTST ↩︎

  253. archimer.ifremer.frarXiv:2207.01372v1 [eess.IV] 4 Jul 2022 - Archimer ↩︎

  254. researchgate.netThe assimilation performance and computational efficiency of ML-4DVAR... - ResearchGate ↩︎

  255. ecmwf.intLinearised physics: the heart of ECMWF's 4D-Var ↩︎

  256. mdpi.comA Framework for Four-Dimensional Variational Data Assimilation Based on Machine Learning - MDPI ↩︎

  257. adapt.psu.eduSystematic Comparison of Four-Dimensional Data Assimilation Methods With and Without the Tangent Linear Model Using Hybrid Backg - PSU ADAPT ↩︎

  258. gmd.copernicus.orgEfficient high-dimensional variational data assimilation with machine-learned reduced-order models - GMD ↩︎

  259. arxiv.org[0909.1678] A localization technique for ensemble Kalman filters - arXiv ↩︎

  260. mdpi.comA Deep Neural Network-Ensemble Adjustment Kalman Filter and Its Application on Strongly Coupled Data Assimilation - MDPI ↩︎

  261. pmc.ncbi.nlm.nih.govAn Integration of Deep Neural Network-Based Extended Kalman Filter (DNN-EKF) Method in Ultra-Wideband (UWB) Localization for Distance Loss Optimization - PubMed Central ↩︎

  262. cisl.ucar.eduConvolutional neural network-based adaptive localization for an ensemble Kalman filter | Computational and Information Systems Lab - cisl.ucar.edu ↩︎

  263. journals.ametsoc.orgOn Domain Localization in Ensemble-Based Kalman Filter Algorithms in - AMS Journals ↩︎

  264. arxiv.orgAccelerate Coastal Ocean Circulation Model with AI Surrogate - arXiv ↩︎

  265. collaborate.princeton.eduA Stable Implementation of a Data-Driven Scale-Aware Mesoscale Parameterization ↩︎

  266. oceans.mit.eduCapturing missing physics in climate model parameterizations using ↩︎

  267. arxiv.orgGeneralizable neural-network parameterization of mesoscale eddies in idealized and global ocean models - arXiv ↩︎

  268. researchgate.net(PDF) Implementation and Evaluation of a Machine Learned Mesoscale Eddy Parameterization Into a Numerical Ocean Circulation Model - ResearchGate ↩︎

  269. zanna-researchteam.github.ioData‐Driven Equation Discovery of Ocean Mesoscale Closures - Computational Oceanography + Climate @ NYU ↩︎

  270. pmc.ncbi.nlm.nih.govPhysics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations - PMC - NIH ↩︎

  271. researchgate.net(PDF) Parameterizing Vertical Mixing Coefficients in the Ocean Surface Boundary Layer Using Neural Networks - ResearchGate ↩︎

  272. mom6.readthedocs.ioVertical Parameterizations - MOM6's documentation! - Read the Docs ↩︎

  273. mom6.readthedocs.ioInternal Vertical Mixing - MOM6's documentation! - Read the Docs ↩︎

  274. cesm.ucar.eduParameterizing Vertical Turbulent Mixing Coefficients In The Ocean Surface Boundary Layer Using Neural Networks ↩︎

  275. journals.ametsoc.orgImplementing Hybrid Background Error Covariance into the LETKF with Attenuation-Based Localization: Experiments with a Simplified AGCM in - AMS Journals ↩︎

  276. repository.library.noaa.govU‐Net Kalman Filter (UNetKF): An Example of Machine Learning‐Assisted Data Assimilation - the NOAA Institutional Repository ↩︎

  277. arxiv.orgOn building the state error covariance from a state estimate - arXiv ↩︎

  278. mdpi.comLength Scale Analyses of Background Error Covariances for EnKF and EnSRF Data Assimilation - MDPI ↩︎

  279. gmao.gsfc.nasa.govBackground Error Covariance Estimation using Information from a Single Model Trajectory with Application to Ocean Data Assimilation into the GEOS-5 - NASA GMAO ↩︎

  280. amazon.sciencePhysics-constrained machine learning for scientific computing - Amazon Science ↩︎

  281. kluedo.ub.rptu.dePhysics-Constrained Deep Learning for Accelerating Climate Modeling - kluedo ↩︎

  282. arxiv.orgHard-Constrained Deep Learning for Climate Downscaling - arXiv ↩︎

  283. cambridge.orgThe challenge of land in a neural network ocean model | Environmental Data Science ↩︎

  284. gmd.copernicus.orgConservation laws in a neural network architecture: enforcing the atom balance of a Julia-based photochemical model (v0.2.0) - GMD ↩︎

  285. arxiv.orgGenerative AI models capture realistic sea-ice evolution from days to decades - arXiv ↩︎

  286. arxiv.orgGenerative AI models enable efficient and physically consistent sea-ice simulations - arXiv ↩︎

  287. mdpi.comSurrogate Model Development for Slope Stability Using Machine Learning - MDPI ↩︎

  288. agu.confex.comMulti-scale, Long-term Stable, and Physically-consistent AI-based Ocean Modeling and Downscaling ↩︎

  289. tc.copernicus.orgData-driven surrogate modeling of high-resolution sea-ice thickness in the Arctic - TC ↩︎

  290. arxiv.orgarxiv.org ↩︎

  291. academic.oup.comPhysics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations | National Science Review | Oxford Academic ↩︎

  292. sp.copernicus.orgCrafting the Future: Machine learning for ocean forecasting - Reports ↩︎

  293. ieeexplore.ieee.orgPhysics Informed Neural Networks for Modeling Large-Scale Wind Driven Ocean Circulation - IEEE Xplore ↩︎

  294. mdpi.comInvestigation of Physics-Informed Methods for Improving Sea Surface Height Prediction Based on Neural Networks in the South China Sea - MDPI ↩︎

  295. arxiv.orgDeep learning in the abyss: a stratified Physics Informed Neural Network for data assimilation - arXiv ↩︎

  296. sp.copernicus.orgData assimilation schemes for ocean forecasting: state of the art - Reports ↩︎

  297. mdpi.comAn Adaptive Variance Adjustment Strategy for a Static Background Error Covariance Matrix—Part I: Verification in the Lorenz-96 Model - MDPI ↩︎

  298. arxiv.orgA unified neural background-error covariance model for midlatitude and tropical atmospheric data assimilation - arXiv ↩︎

  299. journals.ametsoc.orgImproving Variational Data Assimilation through Background and Observation Error Adjustments in - AMS Journals ↩︎

  300. gmao.gsfc.nasa.govBackground Error Covariance Estimation using Information from a Single Model Trajectory with Application to Ocean Data Assimilation into the GEOS-5 - NASA GMAO ↩︎

  301. arxiv.org[2311.02517] A stable implementation of a data-driven scale-aware mesoscale parameterization - arXiv ↩︎

  302. collaborate.princeton.eduA Stable Implementation of a Data-Driven Scale-Aware Mesoscale Parameterization ↩︎

  303. researchgate.net(PDF) Stochastic‐Deep Learning Parameterization of Ocean Momentum Forcing - ResearchGate ↩︎

  304. zanna-researchteam.github.ioData-Driven Equation Discovery of Ocean Mesoscale Closures ↩︎

  305. siam.orgMachine Learning for Multiscale Systems: From Turbulence to Climate Prediction - SIAM.org ↩︎