Current approaches to simulating and predicting the El Niño-Southern Oscillation (ENSO) rely primarily on two classes of models:physics-based dynamical models and data-driven statistical or machine learning models.Physics-based models explicitly represent physical processes and provide mechanistic insight,but their performance is often constrained by model resolution,parameterizations,and limited predictive skill.In contrast,data-driven models excel at capturing complex nonlinear spatiotemporal patterns and achieving high short-term prediction accuracy;however,they frequently lack physical constraints and robust generalization capability.Fusion modeling approaches that combine physical knowledge with artificial intelligence (AI) have therefore emerged as a promising avenue for overcoming these limitations.By embedding AI techniques into dynamical models or incorporating physical constraints into data-driven frameworks,fusion approaches can simultaneously enhance predictive accuracy,physical consistency,and interpretability.Recent studies demonstrate that such integrations improve model parameterizations,strengthen model robustness,and advance ENSO predictability.This paper presents three representative cases illustrating fusion modeling applications in ocean-atmosphere coupled studies.
1) Physics-informed neural-network parameterization of ocean vertical mixing.Uncertainties and biases in ocean vertical-mixing parameterizations are among the primary sources of error in oceanic and climate simulations.Owing to limited understanding of the underlying processes,traditional physics-based parameterization schemes often perform unsatisfactorily in the tropical Pacific,leading to systematic biases in simulations of climatological mean state and ENSO variability.Recent advances in deep-learning methodologies,together with the increasing availability of long-term turbulence observations,provide new opportunities to develop data-driven approaches for parameterizing oceanic vertical-mixing processes.Zhu et al.(2022) introduced a novel parameterization based on an artificial neural network trained using a decadal-long record of hydrographic and turbulence observations in the tropical Pacific.This data-driven parameterization achieves higher accuracy than existing schemes while exhibiting good generalization ability under imposed physical constraints.When integrated into an ocean general circulation model (OCM),the physics-informed neural network (PINN)-based parameterization substantially improves simulations in both ocean-only and coupled modeling configurations.As a novel application of machine learning in physical oceanography and climate science,these activities demonstrate the feasibility of constructing physically consistent,data-driven parametrizations using limited observations and well-established physical constraints to improve climate simulations.
2) U-Net-based representation of surface wind stress anomalies and its integration with an intermediate coupled model for ENSO studies.Numerous dynamical and statistical models have been developed to simulate and predict ENSO.In simplified coupled ocean-atmosphere models,the relationship between sea surface temperature anomalies (SSTAs) and surface wind stress (τ) anomalies is often represented using statistical methods such as singular value decomposition (SVD),which captures only linear responses of wind stress to SSTAs.In recent years,the application of artificial intelligence (AI) to climate modeling has shown considerable promise,and the integration of AI-based models with dynamical models has become an active area of research.Previous studies have demonstrated that AI-based τ models,when trained with extensive datasets,can effectively represent nonlinear relationships among climate variables.
Du and Zhang (2024) developed U-Net-based models to represent the relationship between τ anomalies and SSTAs over the tropical Pacific.The resulting U-Net derived τ model,denoted as τ_UNet,was used to replace the original SVD-based τ model in an intermediate coupled model (ICM),thereby forming a new AI-integrated coupled system,referred to as ICM-UNet.Simulation results from the ICM-UNet indicate that the model can reasonably represent the spatiotemporal variability of oceanic and atmospheric anomaly fields in the equatorial Pacific.A series of experiments were conducted to evaluate model performance.In an ocean-only configuration,the U-Net-derived wind stress anomaly fields were used to force the ocean component of the ICM,yielding realistic simulations of typical ENSO events.These results demonstrate the feasibility of integrating AI-derived model atmospheric components with physics-based dynamical model for ENSO studies.Moreover,the successful coupling of a dynamical ocean model with an AI-based wind stress model provides a novel framework for investigating ocean-atmosphere interactions.
The ICM-UNet simulations reproduce quasi-periodic variations in both atmospheric and oceanic anomaly fields,with the spatiotemporal evolution of SSTAs exhibiting physically consistent patterns.These findings suggest that AI-derived models can serve as effective components within dynamical frameworks for representing ENSO variability.In addition,the ocean component of the IOCAS (Institute of Oceanology, Chinese Academy of Sciences) ICM,when forced by the U-Net-derived wind stress anomalies,is able to reasonably capture typical El Niño events.This case study further confirms the potential of integrating AI-based wind models with dynamical ocean models as a promising approach for hybrid climate modeling.
Nevertheless,the present study represents an initial attempt at such integrations,and several limitations remain.For example,the simulated SSTAs exhibit relatively regular temporal evolution,which does not fully capture the diversity and irregularity of observed ENSO events.As neural networks are inherently data-driven,they can learn nonlinear relationships without physical constraints,whereas the original ICM represents linear relationships based on SVD analysis.Modifying individual components within a coupled system may therefore introduce unintended effects,such as reduced variability in other regions of the Pacific.Furthermore,the potential impacts of the AI-based τ model on other components of the coupled system were not fully assessed.Additional validation is required to ensure that the integrated model adheres to fundamental physical laws.Computational efficiency is another concern,as the AI-integrated model currently incurs higher computational costs.
Future work should focus on several key directions.First,the adaptability of the ICM-UNet should be improved to better capture the diversity of ENSO events,potentially through parameter optimization or the development of more advanced AI-based τ models.Second,comprehensive validation experiments are needed to assess the impacts of the AI-based wind stress model on other components of the coupled system,with adjustments made as necessary to preserve essential ocean-atmosphere dynamical processes.Third,improvements in data exchange efficiency and computational resource utilization are required to enable longer simulations at reduced computational cost.Finally,future studies may explore the construction of alternative AI-based τ models to represent nonlinear interactions among key physical variables and integrate them into other coupled modeling frameworks.This approach holds significant promise as an effective interface between AI techniques and physics-based models in physical oceanography and atmospheric sciences.Ultimately,physics-informed and data-driven integration is expected to establish a new paradigm for ENSO simulation and prediction,with broader implications for climate modeling and sustainable decision-making.
3) A hybrid coupled model for the tropical Pacific based on ROMS and statistical atmospheric models.Coupled climate models with varying levels of complexity often exhibit substantial biases and inter-model differences in simulating ENSO,highlighting the need for alternative strategies,including ICMs,HCMs,and OGCMs.The Regional Ocean Modeling System (ROMS) is a state-of the-art ocean model widely used in regional studies and has been coupled with various atmospheric models.However,its application to basin-scale ENSO simulations in the tropical Pacific has remained largely unexplored.
This study presents,for the first time,the development of a basin-scale hybrid coupled model for the tropical Pacific that integrates ROMS with a statistical atmospheric model representing interannual relationships between SST and surface wind stress anomalies.Two atmospheric wind stress models are implemented:one based on an SVD-derived statistical formulation (denoted as HCM^SVD_OGCM) and the other based on a U-Net-derived AI formulation (denoted as HCM^AI_OGCM).The performance of these two HCM configurations is evaluated in terms of their ability to simulate the annual mean state,seasonal variability,and interannual variations of the tropical Pacific Ocean.
The results demonstrate that both HCMs are capable of reproducing the ENSO cycle,with a dominant oscillation period of approximately two years.Notably,the AI-based configuration,HCM^AI_OGCM,more effectively captures the irregularity and diversity of ENSO events compared with SVD-based configuration,HCM^SVD_OGCM.The ROMS-based HCM developed in this study thus provides an efficient and robust framework for investigating climate variability and ocean-atmosphere interactions in the tropical Pacific.