Abstract. High-resolution reconstruction of ocean dynamics is challenging because spectral bias and the neglect of cross-variable couplings in existing super-resolution (SR) methods often lead to over-smoothed, physically inconsistent outputs, limiting their utility for eddy parameterizations. To overcome these limitations, we present a Multiphysics Super-Resolution version 1.0 (MSR v1.0) framework that jointly reconstructs eight closure-relevant diagnostics—vorticity, deformation measures, stress tensor components, and subgrid momentum forcing terms—directly from low-resolution (LR) velocity fields, consistency is maintained under a filtering scale that aligns with the definition of subgrid tendencies. Our approach integrates three key components: (1) a Dynamic Enhancement Feature (DEF) module to prioritize dynamically active regions; (2) a High-Frequency Enhancement (HFE) module that fuses spatial and spectral operators via learned gating to restore suppressed fine-scale structures such as fronts and eddy rims; and (3) a Physical Consistency Loss that aligns derivative-level structures and algebraic couplings across diagnostics. Experiments on an eddy-resolving simulation dataset across multiple ocean basins and downscaling factors show that MSR consistently outperforms strong SR baselines, yielding sharper reconstructions with improved high-wavenumber spectra and cross-variable consistency. The MSR-reconstructed diagnostics are closure-ready for low-resolution ocean models and can inform or constrain eddy parameterizations, providing a practical, spectrally selective, scale-aware surrogate for high-fidelity multiphysics diagnostics.

To better understand the impacts of airu2013sea feedbacks on simulating tropical cyclones (TCs), this study evaluates the performance of a high-resolution climate model FGOALS-f3-H in simulating TCs over the western North Pacific (WNP), with focus on comparing differences between coupled and atmosphere-only simulations. The TCs are tracked by the TempestExtremes v2.1 framework and classified by minimum sea level pressure. The TC genesis potential index (GPI) is used to analyze TC genesis. The climatology, variability, and environmental factors are evaluated. The results show that both simulations reproduce basic TC characteristics, though notable biases from model deficiencies and airu2013sea coupling remain. The coupled model shows an asymmetric response of surface heat fluxes (SHF) to TC-induced sea surface temperature (SST) cooling depending on TC intensity, which reduces (slightly increases) SHF for intense (weak) TCs. This results in less precipitation, more diffuse warm cores, and weaker winds for intense TCs, but more precipitation for weak TCs. The asymmetric response also produces more realistic TC intensity distributions and windu2013pressure relations (WPR) in the coupled model, although both models overestimate total TC numbers and concentrate in producing medium-intensity TCs. For TC frequency and genesis, both models show profound overestimation in the eastern WNP, and the coupled model has stronger overestimation, which can be explained by the impacts of warm climatological SST biases that enhance TC-favorable environmental factors related with eastward-shifted western Pacific tropical easterly jet (WPTEJ) and weakened subtropical high (WPSH), among which relative humidity and vertical wind shear dominate the GPI biases. Both models cannot capture comparable correlations with El Niu00F1ou2013Southern Oscillation (ENSO), but the atmosphere-only model can capture the correlation with eastern Indian Ocean SST, showing better interannual variability. The coupled model can capture the interannual modulation of the upper ocean heat content (OHC), although it simulates poorer interannual variations and unrealistic TCu2013ENSO response. These findings highlight that airu2013sea coupling corrects TC intensity-related biases but worsens TC locations and variability due to common tropical biases in coupled models.

Submesoscale processes play important roles in vertical heat and mass transport, modulating mesoscale eddies and the energy cycle; thus a parameterization is essential for most ocean models due to submesoscale's spatial scales (-100 m-10 km). This study describes the impact of the submesoscale parameterization scheme by Zhang et al. (2023; Zhang23) in a regional eddy-resolving ocean model in the North Pacific. Compared with the numerical experiment without the scheme, the simulated winter mixed-layer depth (MLD) bias is reduced by 70 % in the Kuroshio Extension (KE) region and the KE jet shifted southward from 36.5 degrees N to 35.5 degrees N, closer to observations. Surface cold biases at 32 degrees-34 degrees N and subsurface warm biases at 36-40 degrees N are reduced by -1 degrees C and -2 degrees C across four seasons, respectively. The effect of submesoscale vertical buoyancy fluxes (VBF) on winter MLD is debated. While widely shown to promote basin-scale shoaling via restratification, they are also known to cause powerful, localized deepening in regions with strong fronts and air-sea interaction. Focusing on this latter scenario, our study reveals a more detailed mechanism, notably distinguishing between local (direct) and remote (indirect) impacts on circulation in the mixed layer and subsurface. Enhanced submesoscale VBF drives weatherscale MLD deepening and subduction along tilted isopycnals in boreal winter in the most active eddy region, mainly limited to 38 degrees-42 degrees N/140 degrees-150 degrees E, promoting southward subsurface cooling and strengthening ocean memory. This feedback modulates the KE's large-scale circulation by shifting its path southward, reducing downstream heat transport, and promoting stratification and shoaling in the eastern region throughout all seasons. These findings demonstrate the importance of submesoscale parameterization for improving simulations of western boundary current systems and highlight its effects in representing remote and subsurface dynamic processes.

The present study compares the Atlantic Meridional Overturning Circulation (AMOC) in the North Atlantic from two simulations by an oceanic general circulation model with 1 degrees x 1 degrees and 0.1 degrees x 0.1 degrees resolution, respectively, which explores the sensitivity of AMOC to the resolution. The ocean model is the latest version of the LASG/IAP Climate System Model (LICOM3), and it is forced with atmospheric data from phase 2 of the Ocean Model Intercomparison Project (OMIP2). Comprehensive comparison between the two simulations indicates that the simulated AMOC is highly sensitive to the spatial resolution. The high-resolution model (LICOM-H) simulates a deeper mixed-layer depth (MLD), due to increased surface salinity, than that in the low-resolution model (LICOM-L) in the Labrador Sea, and it also has stronger AMOC strength (the maximum climatic mean AMOC from 1958 to 2018 is 23.4 Sv for LICOM-H and 21.2 Sv for LICOM-L) and larger variability of AMOC index (2.3 Sv for LICOM-H and 1.7 Sv for LICOM-L). These differences can mainly be attributed to two main dynamic processes that benefit from high resolution. Firstly, LICOM-H can simulate much stronger boundary currents (similar to 0.6 m s-1 ) than LICOM-L (0.3 m s-1 ) in the upper ocean, which leads to saltier and warmer seawater being transported to the Labrador Sea, where it enhances deep convection. Secondly, in LICOM-H, the stronger variability of AMOC is related to the higher sensitivity of the MLD and transformation to the prescribed atmospheric forcing in the Labrador Sea.

Abstract Open water in sea ice significantly influences Arctic low‐cloud formation. However, current global climate models (GCMs), constrained by grid‐scale coupling between the atmosphere and surface components, cannot resolve subgrid heat fluxes and clouds contrasts between open water and sea ice, contributing to biases in Arctic low‐cloud simulations. Here, we develop a novel framework that explicitly resolves subgrid variations in surface component fluxes to the atmosphere in a state‐of‐the‐art GCM. Compared with the conventional framework, the new scheme significantly improves the simulated Arctic low‐cloud seasonality, increasing the correlation coefficient with observations from 0.55 to 0.70. This improvement arises from the seasonal variations in subgrid low‐cloud contrasts between open water and sea ice. The contrast is inversely proportional to the fractional coverage of open water during cold seasons, becoming pronounced when coverage falls below ∼30% at 2° resolution, due to substantial heat and moisture exchanges between small open water and the atmosphere.

The Indonesian Throughflow (ITF) is a key part of the global thermohaline circulation. However, the mechanisms and environmental impacts of the ITF in the geological past remain inadequately investigated. Here, we present model outputs from a set of Last Interglacial (LIG) snapshot simulations carried out by CAS-FGOALS (the Chinese Academy of Sciences Flexible Global Ocean–Atmosphere–Land System model) under the protocol of PMIP (Paleoclimate Modeling Intercomparison Project). Compared to the piControl simulations (the annual mean ITF flux was 18.46 Sv), an annual mean ITF flux increase of about 30.6%–35.9% was found in the early LIG snapshots (24.11–25.08 Sv). This enhancement of the ITF was closely correlated with a La Niña-like state and increased Pacific upwelling. During the LIG, an increased zonal sea surface temperature gradient and intensified easterlies in the tropical Pacific acted as positive feedback loop, resulting in a La Niña-like state with rising sea level in the western Pacific. As a result, the sea level gradient between the tropical western Pacific and the tropical eastern Indian Ocean increased, enhancing the pressure contrast between the two basins and ultimately strengthening the ITF volume transport. Additionally, increased North Atlantic Deep Water formation during the LIG led to enhanced Pacific deep upwelling, which contributed to the strengthening of ITF transport. Comparisons between models and proxies further support a La Niña-like state, enhanced Atlantic Meridional Overturning Circulation, and intensified ITF transport during the early LIG.

To better understand the impacts of air–sea feedbacks on simulating tropical cyclones (TCs), this study evaluates the performance of a high-resolution climate model FGOALS-f3-H in simulating TCs over the western North Pacific (WNP), with focus on comparing differences between coupled and atmosphere-only simulations. The TCs are tracked by the TempestExtremes v2.1 framework and classified by minimum sea level pressure. The TC genesis potential index (GPI) is used to analyze TC genesis. The climatology, variability, and environmental factors are evaluated. The results show that both simulations reproduce basic TC characteristics, though notable biases from model deficiencies and air-sea coupling remain. The coupled model shows an asymmetric response of surface heat fluxes (SHF) to TC-induced sea surface temperature (SST) cooling depending on TC intensity, which reduces (slightly increases) SHF for intense (weak) TCs. This results in less precipitation, more diffuse warm cores, and weaker winds for intense TCs, but more precipitation for weak TCs. The asymmetric response also produces more realistic TC intensity distributions and wind-pressure relations (WPR) in the coupled model, although both models overestimate total TC numbers and concentrate in producing medium-intensity TCs. For TC frequency and genesis, both models show profound overestimation in the eastern WNP, and the coupled model has stronger overestimation, which can be explained by the impacts of warm climatological SST biases that enhance TC-favorable environmental factors related with eastward-shifted western Pacific tropical easterly jet (WPTEJ) and weakened subtropical high (WPSH), among which relative humidity and vertical wind shear dominate the GPI biases. Both models cannot capture comparable correlations with El Niño–Southern Oscillation (ENSO), but the atmosphere-only model can capture the correlation with eastern Indian Ocean SST, showing better interannual variability. The coupled model can capture the interannual modulation of the upper ocean heat content (OHC), although it simulates poorer interannual variations and unrealistic TC-ENSO response. These findings highlight that air–sea coupling corrects TC intensity-related biases but worsens TC locations and variability due to common tropical biases in coupled models.

The El Ni & ntilde;o-Southern Oscillation (ENSO) is the most prominent mode of interannual climate variability; its simulation performance represents a critical benchmark for evaluating the fidelity of coupled climate models. Increasing model resolution is an effective approach to improve the climate model performance; however, the impact of refining horizontal resolution from the hundred-kilometer scale to the tens-of-kilometer scale on ENSO simulation, as well as the underlying mechanisms, remains unclear. This study provides a process-based evaluation of ENSO behavior in two versions of the Chinese Academy of Sciences Flexible Global Ocean-Atmosphere-Land System Finite-Volume version 3 (FGOALS-f3) climate system model: a low-resolution configuration (similar to 100 km; FGOALS-f3-L, hereafter f3-L) and a high-resolution configuration (similar to 25 km; FGOALS-f3-H, hereafter f3-H). Using a reproducible diagnostic framework, we assess how horizontal resolution influences ENSO amplitude, oscillation irregularity, key air-sea coupling processes, and high-frequency (HF) atmospheric variability. The low-resolution version severely overestimates ENSO amplitude, whereas f3-H produces amplitude closer to the observation. Process-based diagnostics show that this improvement arises from the more realistic representation of thermocline and zonal advection feedback processes in f3-H, which arises from the more realistic representation of the meridional structure of ENSO-related zonal wind stress anomalies over equatorial Pacific in f3-H and can be traced back to its improved horizontal resolution. The ENSO cycle in f3-L exhibits excessive regularity, featuring periodic warm-cold transitions; while f3-H reproduces an irregular oscillation resembling the observation. The excessive regularity in f3-L is attributed to its coarser resolution, which limits the simulation performance of tropical cyclones and consequently weakens high-frequency westerly wind activity over the tropical Pacific. The weak stochastic forcing in f3-L is insufficient to disrupt its overly intense ENSO cycle, yielding an overly regular oscillation. By identifying the structural sources of ENSO biases across resolutions, this study provides a reproducible and model-agnostic framework for diagnosing resolution effects on ENSO performance in climate models and informs future development of the FGOALS model family.

Numerical experiments from phase II of the Ocean Model Intercomparison Project (OMIP2) generally underestimate the asymmetric temperature response to interannual changes in the wind‐stress over the equatorial western Pacific subsurface. Despite the knowledge that such a bias may be directly behind the general lack of asymmetry in model‐simulated El Niño‐Southern Oscillation (ENSO), its causes remain unknown. Here we report that a weaker response in the zonal currents to wind forcing in these models results in a biased response in the vertical motion which in turn causes weaker asymmetric temperature responses through the contribution of the vertical motion to the advective heating (dynamic heating). This finding underscores the critical importance of accurately modeling the zonal currents—a previously less noticed element in the modeling effort—to better capture the asymmetric responses in the tropical western Pacific to external wind forcing and thereby improving simulations of ENSO by coupled climate models.

Abstract. Mesoscale Convective Systems (MCSs) are critical components of the climate system and are frequently responsible for extreme precipitation and other catastrophic weather events. Rapid and accurate identification of MCSs can significantly enhance our ability to respond to such extreme events. Traditionally, MCSs identification has relied on threshold-based methods, which are often limited by slower processing speeds and smaller detection areas. Recent advancements in deep learning techniques for object recognition offer a promising alternative for MCSs identification. In this study, we propose an advanced approach to address the challenges associated with traditional threshold-based MCSs identification by creating a specialized dataset and training an MCSs recognition model. First, we constructed an MCSs identification dataset based on infrared satellite data, covering a spatial range (60° S – 60° N, 180° W – 180° E), and a temporal range from 2011 to 2023. Subsequently, by integrating a significance learning strategy and a multi-scale feature extraction method, we developed MCSeg, a novel MCSs recognition model tailored specifically for mid- and low-latitude regions. Finally, we compared the MCSs identified using MCSeg with those identified using the threshold method and conducted precipitation event analysis. The results of the two methods showed a high degree of consistency, indicating the feasibility of applying deep learning methods to MCSs identification.

In this study, we examined the key parameters within deep convection scheme and cloud physics scheme of the CAM6 model to ascertain their significance and influence on simulating mean precipitation in the tropical Pacific. Through simultaneously perturbing 12 selected parameters from deep convection and cloud physics schemes, we conducted perturbed parameter ensemble (PPE) experiments with 128 members. Our analysis uncovered that the parameters showing the most influential effects on tropical Pacific precipitation simulations can be separated into two distinct categories: those primarily governed by the convection scheme, which reflects the competition between convective and large‐scale precipitation, and those predominantly influenced by cloud ice processes. Furthermore, we revealed the importance of nonlinear effects of these perturbed parameters on the simulation of mean precipitation and interpreted the underlying mechanisms. Some biases in simulating precipitation revealed by our PPE experiments align with those in AMIP simulations, offering valuable insights for the AGCM's advancement.

The Indonesian Throughflow (ITF) plays a pivotal role in large-scale ocean-atmosphere interactions in the tropics, regulating the heat and freshwater budget between the Pacific and Indian Oceans. In the context of global warming in the 21st century, The Indonesian Throughflow are projected to be weaken (medium confidence) by CMIP6 simulations. As an analog of possible future warming, the Last Interglacial (LIG, Marine Isotope Stage 5e or Eemian), with global surface temperature reached about 2 °C above present, serves as an outstanding period to explore the climate response to the external forcing and the mechanisms behind it. We use the model outputs from a set of Last Interglacial snapshot simulations carried out by CAS-FGOALS (the Chinese Academy of Sciences Flexible Global Ocean–Atmosphere–Land System model) under the protocol of PMIP for four time periods at 130, 128, 125, and 115 ka. Compared to the piControl simulations (the annual mean ITF flux is 18.46Sv), an annual mean ITF flux increase of about 30.6% - 35.9% was found in the LIG snapshot simulations (24.11 - 25.08Sv). During the LIG, the tropical western Pacific Ocean thermocline was deepened while the tropical eastern Indian Ocean thermocline was relatively shallowed, which was closely tied to the strengthening of the surface easterlies above the tropical western Pacific. Correspondingly, the gradient of the sea surface height between the tropical western Pacific and the tropical eastern Indian Ocean increased, causing pressure contrast between the two basins and probably contribute to the ITF strengthening. We also find that the thermocline gradient between the tropical western Pacific and tropical eastern Pacific was increased, suggesting a La Niña-like state during the LIG. Comparisons of models and proxies further support our conclusions. An examination of the changes in the thermocline water temperature (TWT) record from the eastern Indian Ocean found an enhancement of ITF during MIS 5. Besides, the Maritime Continent was supposed to be more humid by pollen records from west Java and sediment composition from Halmahera Sea. Further analysis suggested that the strengthened ITF during the LIG is inconsistent with the weakened one in the 21st century. While the future global warming is primarily driven by increased CO2 levels, the climate changes during the LIG were principally caused by changes in orbital parameters.

The presence of the Tibetan Plateau is believed to lower pCO2atm by stimulating weathering carbon sink, during which the global ocean is considered a passive carbon reservoir despite the tremendous marine carbon inventory. Yet, recent studies reveal that the orographic forcing of the Tibetan Plateau could lead to drastic changes in ocean circulation, which would substantially affect basin-scale carbon storage and hence pCO2atm. However, this connection between the presence of the Tibetan Plateau and changes in the oceanic carbon inventory remains insufficiently investigated. Here, by employing a state-of-the-art ocean-biogeochemical model, we explore the role of the Tibetan Plateau in determining basin-scale carbon storage patterns based on an idealized experimental design. We find that the presence of the Tibetan Plateau substantially enhances deep Pacific carbon storage and hence lowers pCO2atm via essential reorganization of the meridional overturning circulation, particularly associated with the development of the Pacific halocline. Moreover, the presence of the Tibetan Plateau greatly affects the oceanic carbon uptake in the Northern Hemisphere, which is likely controlled by the variations in surface alkalinity.

Short-term sea surface temperature (SST) forecasting is an essential operational task around China seas. However, the capability of short-term SST forecast from the dynamical numerical model for China seas has not been fully evaluated so far. We assessed the short-term SST forecast skill using a global eddy-resolving ocean forecast system, i.e., the LICOM Forecast System version 1.0 (LFS v1.0) for China seas in 2022 against satellite SST. Results show that LFS v1.0 was able to forecast the short-term SST variation in the study area. The SST with 1-, 7-, and 15-d lead time well captured the observed SST with average pattern correlation coefficient (PCC) of 0.94, 0.93, and 0.92 throughout 2022, the annual mean bias of the forecasted SST of 0.08, −0.16, and −0.33 °C, and the average root mean square error (RMSE) of 0.61, 0.72, and 0.90 °C, respectively. Geographically, the forecast RMSE with 1-d lead time in China seas increased from south to north, and the values were 0.41 °C in South China Sea (SCS) and 1.31 °C in the Bohai Sea (BS). In addition, LFS v1.0 showed better forecast SST abilities in the SCS and East China Sea (ECS) than those in the Yellow Sea and BS. In the ECS and SCS, the forecasted SST was less influenced by the ocean bottom topography due to accurately simulated ocean circulations like Kuroshio. The RMSEs of the SST forecasted by LFS v1.0 displayed seasonal variations, smaller in the area from the middle of boreal August to the middle of boreal December, and larger in boreal late spring and early summer.

Arctic sea ice is an important component of the global climate system and has experienced rapid changes during in the past few decades,the prediction of which is a significant application for climate models.In this study,a Localized Error Subspace Transform Kalman Filter is employed in a coupled climate system model(the Flexible Global Ocean-Atmosphere-Land System Model,version f3-L(FGOALS-f3-L))to assimilate sea-ice concentration(SIC)and sea-ice thickness(SIT)data for melting-season ice predictions.The scheme is applied through the fol-lowing steps:(1)initialization for generating initial ensembles;(2)analysis for assimilating observed data;(3)adoption for dividing ice states into five thickness categories;(4)forecast for evolving the model;(5)resam-pling for updating model uncertainties.Several experiments were conducted to examine its results and impacts.Compared with the control experiment,the continuous assimilation experiments(CTNs)indicate assimilations improve model SICs and SITs persistently and generate realistic initials.Assimilating SIC+SIT data better corrects overestimated model SITs spatially than when only assimilating SIC data.The continuous assimilation restart ex-periments indicate the initials from the CTNs correct the overestimated marginal SICs and overall SITs remarkably well,as well as the cold biases in the oceanic and atmospheric models.The initials with SIC+SIT assimilated show more reasonable spatial improvements.Nevertheless,the SICs in the central Arctic undergo abnormal summer reductions,which is probably because overestimated SITs are reduced in the initials but the strong seasonal cycle(summer melting)biases are unchanged.Therefore,since systematic biases are complicated in a coupled system,for FGOALS-f3-L to make better ice predictions,oceanic and atmospheric assimilations are expected required.

How much heat is pumped into the interior of the oceans directly affects projected future warmings of the atmosphere and surface climate, with both global and regional implications. Kuroshio-Oyashio Extension region (KOE) influences the local marine ecosystems and is part of the North Pacific decadal variation systems, it also tele-connects with the North American weather and is a key projection indicator for the marine heatwaves. For a more reliable understanding and future projection of the future climate and extreme events in the North Pacific, it is important to predict potential future spatial and temporal warming patterns in KOE more accurately. The future KOE warming pattern and warming mechanisms are analysed, with future scenario simulation by skillful high-resolution coupled model FGOALS-f3-H, compared with low-resolution model FGOALS-f3-L. Results show that high-resolution models simulate a future deep, strong warming reaching 600 m in the Kuroshio-Oyashio region, while the warming in low-resolution models is only above 300 m. Deep warming includes two spatial parts, one in the south of Kuroshio, which is contributed by heaving, and one in the north around Oyashio which is contributed by the northward movement of the subtropical gyre. The skill of high-resolution models to simulate future deep warming is co-contribute by the improvements in the ability to realistically capture the Kuroshio Extension current, with its meridional position, sharp front as well as large horizontal current speed, and mixed layer depth with mesoscale eddies effects.

The El Niño-Southern Oscillation (ENSO) is one of the most significant integrated interannual oscillations with coupled atmosphere-ocean processes in the tropical Pacific. Most coupled climate models are weak in depicting ENSO asymmetry over equatorial Pacific subsurface. And it is still unclear how the stand-alone ocean model contributes to this bias. In this study, we found that most ocean models from the Ocean Model Intercomparison Project (OMIP), driven by JRA55, underestimate the asymmetry of ENSO in the equatorial western Pacific subsurface. We investigated the primary factors contributing to this bias using composite analysis and diagnostics, and found that the weaker responses in upwelling and stronger responses in downwelling to westerly and easterly wind stress anomalies in the models are mainly responsible for the bias. Furthermore, the underestimation of zonal current variability over western Pacific subsurface, influenced by the gradient of mean state of sea surface height along the equatorial Pacific, leads to an opposite relationship between asymmetry and the zonal component of nonlinear dynamic heating in the western Pacific subsurface comparing to that in the eastern Pacific subsurface. Our study emphasizes the importance of accurately modeling ocean currents to capture the characteristics of ENSO nonlinearity and highlights the significance of nonlinear dynamic responses to external forcing.

This paper provides a comparative analysis of the performance of a high-resolution regional ocean-atmosphere coupled model in predicting tropical cyclone(TC)gales over the northern South China Sea.The atmosphere and ocean components of the coupled system are represented by the China Meteorological Administration's Tropical Regional Atmosphere Model for the South China Sea(CMA-TRAMS)and the LASG/IAP Climate system Ocean Model(LICOM),respectively.The Ocean Atmosphere Sea Ice Soil VersionH 3(OASIS3)software has been utilized for the exchange of momentum,heat,and freshwater fluxes between these two components.An assessment of the coupled model's three-day predictions for five TCs'gales was conducted.Preliminary findings indicate that the predicted TC tracks show less sensitivity to oceanic influences than the predicted TC intensities.Significant improvement in predicting the surface TC gales has been achieved through coupling the ocean model.This improvement is attributed to the impact of the warmer ocean's effect on TC intensification,counteracting the cooling effect of the cold wake.In summary,coupling has enhanced the model's predictive capabilities for TC gales.A detailed assessment of the coupled model's performance in predicting other tropical weather phenomena is forthcoming.

Tropical Instability Waves (TIWs) play a crucial role in modulating Sea Surface Temperature (SST) variability in tropical oceans, yet their representation in current forecast systems remains challenging. This study investigates the relationship between TIWs and sub-seasonal SST predictability while evaluating the performance limitations of the Licoms Forecast System. Through comprehensive analysis of observational data and model outputs, we demonstrate that TIWs provide significant potential for enhancing sub-seasonal SST forecast skill through their regular wave patterns and predictable evolution characteristics. However, our findings reveal that the current Licoms forecast systems systematically underestimate both TIW intensity and wavelength. Critical examination of error sources indicates that these deficiencies primarily originate from initialization fields rather than model physics or dynamics.

Existing traditional ocean vertical-mixing schemes are empirically developed without a thorough understanding of the physical processes involved, resulting in a discrepancy between the parameterization and forecast results. The uncertainty in ocean-mixing parameterization is primarily responsible for the bias in ocean models. Benefiting from deep-learning technology, we design the Adaptive Fully Connected Module with an Inception module as the baseline to minimize bias. It adaptively extracts the best features through fully connected layers with different widths, and better learns the nonlinear relationship between input variables and parameterization fields. Moreover, to obtain more accurate results, we impose KPP (K-Profile Parameterization) and PP (Pacanowski–Philander) schemes as physical constraints to make the network parameterization process follow the basic physical laws more closely. Since model data are calculated with human experience, lacking some unknown physical processes, which may differ from the actual data, we use a decade-long time record of hydrological and turbulence observations in the tropical Pacific Ocean as training data. Combining physical constraints and a nonlinear activation function, our method catches its nonlinear change and better adapts to the ocean-mixing parameterization process. The use of physical constraints can improve the final results.