The study utilized version 1.2.1 of the open-source global coupled model CESM, developed by the National Center for Atmospheric Research (NCAR). This integrated model includes the Community Atmospheric Model version 4 (CAM4) , the Parallel Ocean Program version 2 (POP2) , the Community Ice Code version 4 (CICE4), and the Community Land Model version 4 (CLM4), as well as other modules. The atmospheric component has a horizontal resolution of 0.9°×1.25° with 26 vertical levels, while the ocean component was integrated at a nominal resolution of 1° with an enhanced meridional resolution of 0.5° in the equatorial region and 60 vertical levels.

In many OGCMs, vertical mixing can be parameterized separately by region, including upper-boundary-layer schemes and a diapycnal-mixing scheme for the ocean interior. The K-profile parameterization is widely used to parameterize vertical mixing in ocean models. It includes a background diffusivity parameter that determines the diapycnal mixing in the thermocline. It is critical to the heat transfer between the upper boundary layer and the ocean interior. The background diffusivity parameter is typically set to a constant value, and its magnitude is determined by fitting the model to observations or theoretical considerations. As identified by much of the previous work, the background diffusivity parameterization is a key factor in vertical mixing parameterizations, and it has significant uncertainties and contributes to a large bias in SST simulations . have shown that a better background diffusivity parameterization leads to more realistic simulations of the cold tongue and equatorial thermocline, which has the potential to affect the fidelity of simulated seasonal to interannual variability in the tropical Pacific, such as the ENSO phenomenon. Therefore, the present study focused on estimating the parameters in background diffusivity parameterization.

The ocean model of CESM, POP2, was initially proposed by to solve three-dimensional ocean dynamic primitive equations on a global grid under the assumptions of Boussinesq and hydrostatic approximation. The background diffusivity parameter, denoted as kw, is mainly used to characterize mixing processes resulting from the breaking of inertial internal waves . However, due to the uncertain propagation and dissipation behavior of these waves, the parameter value of kw has significant uncertainty. first estimated the averaged diapycnal diffusivity of 10-4 m2 s-1 based on the advective–diffusive balance, thus requiring a background diffusivity of O(10-4 m2 s-1) to realistically produce the pycnocline in numerical models . However, microstructure measurements generally give estimates that can be reduced by 1 order of magnitude . Hence, a constant background diffusivity of O(10-5 m2 s-1) has typically been applied in many ocean and climate modeling studies. Recently, observational evidence has indicated that the assumed constant background diffusivity is not uniform but rather is spatially varying . In particular, microstructure measurements suggest that background diffusivity should be reduced near the Equator, with a magnitude of O(10-6 m2 s-1) . adopted a latitudinal structure of kw in the POP2 model and found that it simulated the ocean state better, which has been subsequently accepted by follow-up studies. Specifically, POP2 utilizes four independent background vertical diffusivity coefficients (BVDCs) to simulate the latitudinal structure of the background diffusivity. Table 1 lists each coefficient's default values and descriptions .

The BVDCs have a total of four constants, comprising a coefficient v1 describing global diffusivity, a coefficient ve describing equatorial diffusivity , a coefficient vp describing the diffusivity near latitudes of 28.9° S and 28.9° N, and a coefficient vb describing diffusivity in the Banda Sea region alone . Among them, vp is also known as the maximum PSI-induced diffusivity, representing the result of the parametric subharmonic instability (PSI) of the M2 tide . Therefore, in POP2, the background diffusivity parameter kw has only a fixed form of spatial variation . Except for the Banda Sea, which takes on a specific value, globally, kw varies only latitudinally and can be expressed as follows: 1kw=ve+v1θ102+vpe-0.4(θ+28.9)+vpe-0.4(θ-28.9), where θ is the latitude.

Therefore, the kw value has been modified from its typically constant value of 0.1 to 0.17cm2 s-1 (v1+ve=0.17) almost everywhere in POP2, and there are regions where different values are used: 1.0 cm2 s-1 in the Banda Sea (vb), 0.3 cm2 s-1 in the latitude bands around 28.9° N and S (v1+ve+vp), and 0.01 cm2 s-1 at the Equator (ve).

The Data Assimilation Research Testbed (DART) was employed in this study to implement the data assimilation system . DART, an open-source software, offers various filter methods' implementations. Previously, this data assimilation system was used to study the impact of initial-state errors on assimilation quality by assimilating ocean observations within a quasi-weakly coupled data assimilation framework . The EAKF method yielded the analysis ensemble, serving as the initial condition for climate variability forecasting. Notably, the system's initial conditions facilitated the forecasting of significant climate variability, such as ENSO and IOD (the Indian Ocean dipole), and consequently directed a demonstrable improvement in forecasting skill .

The description of the assimilation system is presented in detail by . In brief, it is a weakly coupled data assimilation system since only ocean observations are assimilated, and the coupled model is used for integration. This study utilizes the ensemble adjustment Kalman filter approach with 20 ensemble members. The ensemble members are constructed using long-term spin-up integration results and then repeatedly assimilating the WOA18 climatology data over 4 years to correct the climatological bias. This approach is essential to ensure that the initial ensemble can effectively incorporate all observations during the data assimilation procedure.

Two sets of observation data are assimilated every 10 d. One dataset is the Optimum Interpolation Sea Surface Temperature (OISST) dataset version 2.1 retrieved from the National Oceanic and Atmospheric Administration (NOAA). The other is the EN4 profile dataset version 4.2.1 of the UK Met Office. The OISST dataset has a daily 0.25° resolution and was constructed by combining observations from different platforms (satellites, ships, buoys, and Argo floats) on a regular global grid. The EN4 profile dataset is a collection of ocean temperature and salinity (T–S) profiles obtained across global oceans from 1900 to the present. Quality control methods ensure good quality .

The datasets were pre-processed before being assimilated into the system. Regarding the data assimilation system that assimilates SST and T–S profiles every 10 d, daily profiles were merged and assigned to the final day of each sequence. To prevent overfitting due to assimilating excessive profile observations, the data at different depths were first interpolated to 31 layers from 5 m to approximately 2100 m and then averaged horizontally. Specific vertical depths were obtained from the EN4 analysis data . The mean value of all the data in each 1°×1° cell at each level was regarded to be the observation value. Moreover, the OISST data were thinned such that only data on the 1°×1° grids were assimilated every 10 d. Previous studies have shown that this processing method can produce effective state estimation results .

Localization was employed using the Gaspari and Cohn function , which employed a cutoff half-width of 0.1 rad (approximately 600 km) for both observations. The SST and T–S profiles had vertical localization half-widths of 250 and 1000 m, respectively. Additionally, the application of covariance inflation involved utilizing a constant inflation factor with α=1.02 for model states. These factors were determined empirically and verified in prior studies .

One approach to achieving PE is the state vector augmentation method, in which parameters are treated as specific model variables and included in the state vectors. By updating the augmented state vector with observations, the model state and parameters can be estimated concurrently . Applying PE in a CGCM presents several technical challenges. Firstly, many parameters in the GCM emanate from simplifying underlying physical processes, which may display globally uniform values. Updating a few global parameters with numerous data may lead to an accumulation of sample errors, leading to PE failure. Since the parameters estimated in this paper are four constants, this is the main challenge in the experiments. To overcome this hurdle, we used the adaptive spatial averaging (ASA) method designed by for the CGCM. In each data assimilation step, we transformed each parameter from a single scalar value into a two-dimensional field, considering spatial dependence and localization during the assimilation. Afterward, we use the adaptive algorithm to average the two-dimensional parameter fields to produce a scalar value incorporated into subsequent model integration. This algorithm calculates the ratio of the a posteriori standard deviation to the a priori standard deviation at each grid point after each update of the two-dimensional parameters, which implies the strength of the effect of assimilation, and then averages the parameter values at grid points where the ratio exceeds a certain threshold. This threshold is chosen using an adaptive algorithm to ensure that a certain number of grid points (in this experiment, 10 000 out of a total of 80 000 grids) are included in the calculation of the averaged parameters. For more details, refer to .

A further challenge arises from covariance inflation. Studies have noted that the parameter ensemble's spread (standard deviation) is generally relatively lower than that of the state ensemble, primarily because parameters remain constant for the mode integration. Consequently, ensemble-Kalman-filter-based PE requires a larger covariance inflation factor for the parameter ensemble. In a previous study, we employed twin experiments to demonstrate the necessity of covariance inflation for PE of BVDCs in CESM and developed a two-stage covariance inflation approach . Specifically, the conventional covariance inflation was applied to the augmented vector of the model states and the 2-D parameter fields using a fixed inflation factor of α=1.02 before assimilation. Afterward, we average the analysis data of 2-D parameter fields to obtain global scalars and utilize a covariance inflation factor of αp=1.25 solely for the parameter ensemble. This factor is deduced by calculating the average growth rate of the state variables in the model integration. Figure  provides a schematic diagram illustrating the PE process described above.

However, the conditional covariance inflation (CCI) method is usually used in practical data assimilation to ensure that the ensemble spread does not fall below a lower bound . The CCI is designed to inflate the parameter ensemble spread back to a predefined threshold value when it is smaller than the threshold. In this work, we intentionally do not use the CCI method and let the parameter ensemble degenerate after several data assimilation cycles. After that, all ensemble members use the same improved parameters that no longer change value with the subsequent data assimilation. This strategy allows the parameter ensemble to converge and makes subsequent ensemble forecasting experiments easier to implement.

It is also worth noting that, in order to avoid unphysical parameter values, after each parameter estimation, if an abnormal parameter value (e.g., negative value) occurs for an ensemble member, we remove the parameter and use the parameter of the neighboring member to integrate the model.

Conducting sensitivity analyses before PE is necessary to ensure that the parameters have a significant impact on the observed variables. In this study, the sensitivity experiment was initially conducted to show the sensitivity of model temperature and salinity to the BVDCs. An ensemble of size 20 was integrated using the same initial states but with perturbed parameters. We perturbed each BVDC by adding noise generated from a Gaussian distribution with a mean value of 0 and a standard deviation of 30 % of its default value. We measured the variable sensitivity to the perturbed parameters by using the ensemble spread of each variable.

Subsequently, we conducted separate SE and PE experiments using the initial ensemble introduced earlier. The assimilation time window started in January 2005 and continued until December 2017. In the SE experiment, the SST and T–S profiles were assimilated every 10 d to update the model state variables that include temperature, salinity, sea surface pressure, and surface current velocity. The BVDCs listed in Table  were used in model integration during the entire period. The PE experiment used the same observations to update the model state variables and the BVDCs concurrently. As showed, the signal-to-noise ratio of the state–parameter error covariance in the coupled model can be significantly improved after the state estimation reaches quasi-equilibrium. Therefore, we performed only pure state estimation in the first year of the PE experiment and activated the PE function from the beginning of the second year. That is, the parameter values changed gradually from 2006 onwards. At this point, the observation-constrained states can improve the parameter estimates more effectively.

We compared the results of SE and PE experiments with validation data to demonstrate the impact of PE on reducing analysis errors. The temperature and salinity from the objective analysis data of EN4.2.1 are used for validation. It should be noted that the EN4 profile dataset for assimilation is a collection of profiles, and the EN4 objective analysis dataset is processed and gridded data. To ensure impartiality in the validation data, we also incorporated high-quality reanalysis products such as the Ocean Reanalysis System 4 (ORAS4) by and the Geophysical Fluid Dynamics Laboratory's ensemble coupled data assimilation (GFDL/ECDA) by .

The EAKF can provide initial conditions for ensemble prediction by running an ensemble of members. The analysis ensembles of SE and PE experiments were utilized as initial conditions for climate forecasting with the coupled model. We conducted ensemble forecast experiments from 2008 to 2017 using the analysis ensembles derived from both SE and PE. The parameters obtained by PE were also employed in the latter case. Predictions were issued at the beginning of each January, April, July, and October, extending for 12 months. The Hadley Centre sea ice and sea surface temperature dataset (HadISST) served as a reference dataset to compare the produced prediction products.

The schematics in Fig. a–c show the sensitivity experiment, the SE experiment, and the PE experiment, respectively. It can be seen that the sensitivity experiment is a free-integration experiment using the same initial condition and different parameters. The SE experiment uses the ensemble of state variables and the same default parameters. At the same time, the PE experiment uses ensembles for both state variables and parameters. PE experiments are divided into three phases, which we will specify in the “Results and discussions” section. Moreover, it also shows that the state and parameter estimation results are used in the later hindcast experiments.

Twenty identical ensemble members were utilized for the parameter sensitivity experiment in the CESM integration, which persisted for 5 years, with four parameters perturbed simultaneously. It is worth noting that we tried experiments where we perturbed the parameters one by one, and the experimental results showed that perturbing the different BVDCs had a comparable effect (not shown here), so we used this simultaneous perturbation scheme. The ensemble spreads of temperature and salinity variables, which measure their sensitivity to the perturbed parameters, are shown in Fig. 2. The global (66.5° S–66.5° N) and equatorial (5° S–5° N) averaged temperature and salinity ensemble spreads were demonstrated accordingly. Perturbing BVDCs in the model leads to a rapid increase in temperature and salinity ensemble spread within the first year, followed by relative stability in the succeeding years. Figure  shows that temperature variables have the maximum sensitivity to BVDCs at approximately 100 m depth, with salinity variables being most sensitive to these parameters at the sea surface. The influence of parameter uncertainty can extend up to a depth of approximately 400 m. Additionally, the equatorial area is highly sensitive to BVDC parameters in terms of temperature at a depth of 50–100 m and in terms of salinity at the sea surface. The surface temperature variability in Fig. a shows a conspicuous seasonal cycle which can possibly be related to the diverse rates of change in mean temperature instigated by distinct ocean areas between the Northern Hemisphere and the Southern Hemisphere.

The last 3 years' outcomes were used to compute the mean spread and analyze its spatial distribution. Figure  provides additional validation that temperature variability is highest within the equatorial range and most pronounced at a depth of 100 m. In deeper layers, the parameters affect the temperature more significantly in the western Pacific. Additionally, salinity is highly sensitive to the parameters in the warm-pool region of the tropical western Pacific, and the sensitivity of salinity to parameters is highest in the shallow layers of less than 50 m depth. Furthermore, in the extratropics, the temperature and salinity in the Kuroshio Extension and Gulf Stream regions are also sensitive to these parameters to some extent.

The sensitivity experiment shows that the model temperature and salinity are sensitive to the uncertainty in the BVDCs, strongly indicating that assimilating SST and T–S profiles can potentially reduce the uncertainty.

We conducted separate SE and PE experiments, assimilating observations during the period between January 2005 and December 2017. In the SE experiment, default values of the BVDCs were consistently used in all ensemble members throughout the entire period. However, the PE experiment comprised three distinct phases. During the initial phase, we utilized perturbed parameters to perform state estimation. It spans a period of 1 year and brings the state estimation process to approximately quasi-equilibrium, where the uncertainty of coupled model states is sufficiently constrained by observations. In the second phase, spanning from 2006 to 2007, we activated the PE function illustrated in Fig. . This function facilitated continuous correction of the parameter ensemble through observations. Finally, during the third phase, spanning from 2008 to 2017, these parameters remained unchanged.

Figure depicts a graphical representation of the 20 ensemble members of the four BVDCs over time, with the ensemble mean represented by the solid red line. Observations gradually decreased the spread of the parameter ensemble, resulting in less uncertainty. After approximately 2 years, the parameter ensemble degenerated, and the spread reached 0. Consequently, assimilating observations could no longer adjust the parameters.

Table presents the final values of BVDCs. Notably, v1 and vp values are 20 % higher than the default values, while vb is 10 % higher (except for ve, which is slightly lower than the default value). It is also worth noting the almost globally increasing value of the background vertical diffusivity, kw, as calculated through Eq. (1) and depicted in Fig. . The left-hand side of Fig.  displays the band structure of the default background diffusivity, while the increment obtained by PE is shown on the right-hand side, further validating the achieved results.

As previously mentioned, the parameters have remained unchanged since 2008. Consequently, the third phase of the PE experiment can be considered to be a distinct SE experiment using the estimated parameters listed in Table . This study focuses specifically on evaluating the analyses obtained from the third phase by comparing the results of the PE and SE experiments.

We compare our analysis fields to the gridded objective analysis data from EN4 and other reanalysis products to demonstrate the validity of our results. Figure  displays the root mean square error (RMSE) of the temperature in the analysis fields for the period of 2008–2017 by region. We compared the results to EN4, ORAS4, and ECDA. The regions are global (within 66.5° N–S), Pacific, Indian Ocean, Atlantic, and intra-tropical (within 30° N–S). Similar findings can be observed globally and in most regions using different datasets. When examining the global mean temperature, the depths with significant analysis errors are consistent with the parameter-sensitive depths, indicating that parameter uncertainty can impact the analysis accuracy. Moreover, the reduced RMSEs of the PE experiment indicate that PE improves the quality of the analysis. In particular, noticeable improvements are observed below a depth of 100 m. The most pronounced improvement is observed in the Atlantic Ocean and in tropical regions. Figure  illustrates the salinity errors in the analysis. The highest error is observed in the sea surface layer, which is consistent with the depth most sensitive to parameters (Fig. b and d). In contrast to temperature, PE primarily enhances salinity accuracy in deep-Atlantic and tropical regions.

Figure displays the RMSEs of the SE experiment and EN4 data in the tropics while emphasizing the disparity compared to the PE experiment. As Fig. a shows, the most considerable temperature errors appear in all oceans around the depth of 100 m, which matches the sensitivity analysis result for temperature depicted in Fig. c. Figure e denotes an improvement in PE for these errors, implying its usefulness throughout the tropics. It is noted that there is improvement in areas where the temperature is sensitive to the uncertainty of those parameters. The temperature error in the deeper layers of the tropical oceans has also been reduced, especially in the deeper Atlantic Ocean. Figure b, d, and f show the results for salinity. Although not as significant as temperature, the depths and areas where salinity errors emerge in the SE analysis align with those sensitive to parameters. Unsurprisingly, PE partially mitigates these errors, most significantly around the Andaman Sea, in waters near Indonesia, and in the coastal West African waters where vertical mixing is intensive.

The model bias of CESM is relatively large in the Atlantic Ocean. have shown the zonal-mean temperature and salinity of CCSM4 (which uses the same ocean model as CESM) minus climatology from observations. They noted that the deep Atlantic Ocean remains generally warmer than observed by about 0.58 °C in terms of the mean. The local temperature and salinity maxima between 20 and 30° N at a depth of about 1000 m are associated with the warmer- and saltier-than-observed Mediterranean outflow through the Strait of Gibraltar. The largest salty biases occur in the deep Atlantic Ocean. The upper-ocean Atlantic north of 15° N remains mostly saltier than the climatology. We show similar results in Fig. , which displays the zonal-mean temperature and salinity of the Atlantic minus the climatology calculated from EN4 data using SE results and PE results, respectively. By comparing the climatology bias of the results from two experiments, it is seen that the most significant improvement in the PE in relation to the SE lies in the Atlantic Ocean at a depth of 1000 m between 20 and 30° N. It strongly suggests the contribution of improved background diffusivity parameters to reducing model systematic biases. It can be inferred from the conclusions of that this improvement also stems from the improvement in the outward flow in the Strait of Gibraltar. Although the sensitivity and analytical errors in this region cannot be demonstrated directly in Fig.  due to resolution, the effect of PE is demonstrated by affecting the 1000 m Atlantic Ocean between 20 and 30° N. This explains the smaller deep-Atlantic RMSE in the PE results presented in Figs.  and .

This study utilized analysis ensembles from the coupled data assimilation system to conduct ENSO forecast experiments between 2008 and 2017 (as shown in the schematic diagram in Fig. b and c). The Niño 3.4 index, calculated as the averaged sea surface temperature anomalies between the latitudes of 5° S to 5° N and longitudes of 190° E to 120° W, was employed to illustrate the variability of ENSO. The Niño 3.4 indices of the SE and PE forecasts were computed against various lead times. The anomaly correlation coefficients (ACCs) of these outcomes with the index derived from HadISST data were employed to measure the prediction skills, as shown in Fig. a. Moreover, Fig. b depicts the RMSE of the forecasts against HadISST.

The prediction skills of both the SE and PE cases were significantly superior to those of the persistence skills represented by a dotted black line. For lead times exceeding 5 months, the PE case exhibited higher ACCs compared to the SE case. By setting the threshold value of an effective prediction as a ACC of 0.5, which is equivalent to the 99 % statistical confidence level with an independent sample size of 30, it was observed that the SE case effectively predicted ENSO at a lead time of up to 9 months, which is 1 month in advance compared to the PE case. To demonstrate the significance of the PE advantage, ACCs were computed for each ensemble member using HadISST. The shaded areas represent the ACCs of the ensemble mean plus/minus the standard deviation of the ACCs of each member, further confirming the superior prediction ability of the PE case.

The root mean square errors (RMSEs) of the PE case were also lower than those of the SE case, particularly after a lead time of 5 months. Additionally, the ensemble spreads (dashed colorful lines in Fig. b) of the PE results were larger compared to those of the SE predictions. Since the spreads of the PE results were closer to the RMSE than those of the SE results, it is clear that the PE initial conditions are more consistent.

Figure illustrates the spatial correlation coefficient pattern between the predicted sea surface temperature (SST) anomaly and the corresponding HadISST data over the tropical Pacific for the SE and PE cases. The SE and PE results showed no significant difference for lead times of 1 or 4 months. However, for longer lead times, the initial conditions and parameters obtained through PE significantly improved the SST anomaly in the tropical Pacific. As illustrated in Fig. , the outcomes of our data assimilation experiments reveal a notable reduction in analysis error within the thermocline of the equatorial Pacific due to parameter estimation. This yields improved initial conditions for ensemble prediction. Simultaneously, the refined parameters contribute to enhanced global vertical mixing. Consequently, the SSTA in the tropical Pacific, derived from our forecasts incorporating parameter estimation, exhibit a higher ACC with observational data. In addition, since the advantage of the PE analysis in the subsurface layer takes some time to affect the SST, the forecast skill based on the PE results is only significantly greater than that of the SE results at longer lead times.

The improvement in PE for ENSO forecasts is likely attributable to better simulations of subsurface temperatures. Previous studies have consistently shown that the accuracy of initial subsurface conditions is crucial for ENSO event prediction . The PE method not only enhances the accuracy of the initial subsurface conditions (as demonstrated in Figs.  and ) but also provides constrained parameters that more accurately represent the background diffusivity process in the ocean model, leading to improved forecast skills.