Seasonal-to-decadal prediction of El Niño-Southern Oscillation and Pacific Decadal Oscillation

Near-term climate prediction, encompassing seasonal-to-decadal (S2D) timescales, is increasingly important for climate risk management. These predictions are influenced by both boundary conditions (greenhouse gases and aerosols) and initial conditions (ocean state). Initialized prediction systems have shown promise in improving S2D prediction skill, with multi-model ensembles yielding more reliable results than uninitialized projections. Low-frequency sea surface temperature (SST) variability, such as ENSO, PDO, and AMO, is a primary source of S2D prediction skill. While studies have examined S2D prediction of SST variability in various ocean basins, particularly the North Atlantic, a comprehensive assessment using the latest decadal hindcasts for ENSO and PDO in the Pacific is lacking. ENSO prediction has seen advancements, with dynamical models exhibiting high skill up to 12 months, and machine learning techniques extending this to 18 months. However, multi-year ENSO prediction remains less explored, typically focusing on extreme events or single forecasting systems. The PDO, a leading mode of decadal SST variability in the North Pacific, is another key area. Its prediction skill, though reported to extend several years, has been evaluated using limited models due to computational demands. This study addresses these gaps by using a very large ensemble of CMIP5 and CMIP6 retrospective decadal predictions, spanning over half a century, to assess S2D prediction of ENSO and PDO. The study aims to identify the contributions of model initialization and external radiative forcing to prediction skill, and to evaluate the relative importance of ensemble size and multi-model ensemble averaging.

The existing literature highlights the growing importance of skillful near-term climate prediction and the role of both initial and boundary conditions. Studies using CMIP5 and CMIP6 data have shown progress in decadal prediction, particularly for the North Atlantic. However, the Pacific basin, characterized by ENSO and PDO, requires further investigation. While considerable work exists on ENSO prediction (up to 1-1.5 years), multi-year predictions are limited, often focusing on extreme events or single models. Similarly, PDO prediction skill, while demonstrated in some studies, needs assessment using the latest CMIP6 hindcasts and large ensembles to better understand its predictability and the contribution of different factors. Previous research suggests the importance of model initialization and the accurate representation of external radiative forcing for better prediction skill, but the relative contributions of these factors for ENSO and PDO remain unclear.

This study utilized retrospective decadal predictions (hindcasts) from six CMIP5 and ten CMIP6 models, encompassing a total of 142 ensemble members. Data were interpolated to a 2.5° × 2.5° grid, and model drift was addressed through bias correction, subtracting the difference between modeled and observed climatology at each lead time. The multi-model ensemble (MME) was calculated by averaging the bias-corrected ensemble members with equal weighting. Prediction skill was assessed by comparing the MME with observed SST anomalies (SSTA) from NOAA ERSSTv5. ENSO prediction skill was evaluated using the three-month running mean NINO3.4 index, while PDO prediction skill was evaluated using the annual-mean PDO index, computed by projecting model anomalies onto the observed PDO pattern. To quantify skill originating from external radiative forcing, the MME from uninitialized historical simulations was also analyzed. The anomaly correlation coefficient (ACC), mean-squared skill score (MSSS), and ratio of predictable components (RPC) were used to quantify prediction skill and its spread. Statistical significance was determined using a non-parametric bootstrap resampling method. The sensitivity of prediction skill to ensemble size was evaluated by calculating ACCs as a function of ensemble sizes using bootstrap resampling. The theoretical estimation of the ensemble-mean MSE was calculated to quantify the impact of ensemble averaging on prediction skill, assuming perfect model conditions. The study compared practical MSSS with theoretically estimated MSSS to assess the multi-model ensemble average effect. CMIP5 and CMIP6 results were compared to investigate improvements in prediction skills across CMIP phases.

The study found that the overall skill scores of MME SSTA decreased inhomogeneously as lead time increased. Model initialization yielded successful SSTA predictions up to two years, especially in the tropical Pacific. Skillful prediction beyond three years was primarily attributed to external radiative forcing. For ENSO, significant skill was observed at a one-year lead time for amplitude and two years for phase. MME skill scores were generally higher than individual models, particularly in spring and summer, suggesting a multi-model ensemble average effect. Around 40 ensemble members were sufficient for skillful one-year ENSO prediction, while multi-year predictions required more than 70. The improved MME ENSO prediction skill in spring and summer was partly due to the cancellation of model errors. For PDO, prediction skill was more limited than for ENSO. The MSSS for PDO closely followed the theoretical estimation, suggesting that individual model errors were not effectively eliminated by multi-model ensemble averaging. More than 50 ensemble members were needed to achieve maximum prediction skill at YR5–9, although the skill was not very high. CMIP6 showed slightly improved springtime ENSO prediction skill compared to CMIP5, likely due to a larger ensemble size. However, PDO prediction skill was comparable between CMIP5 and CMIP6 at longer lead times, indicating a lack of major improvement.

The findings demonstrate that both model initialization and accurate representation of external radiative forcing are crucial for improved S2D prediction in the Pacific. The higher MME ENSO prediction skill, especially in spring and summer, highlights the benefit of multi-model ensemble averaging in mitigating the spring predictability barrier. The limited PDO prediction skill, compared to ENSO and AMO, suggests challenges in representing PDO-related climate variability in current decadal prediction systems. The study suggests that advanced post-processing techniques or statistical error correction might enhance PDO prediction. The similarity between practical and theoretical MSSS for PDO hints at common errors among models in PDO prediction. The indistinguishable resemblance between ENSO and PDO and systematic model bias in the extratropical Pacific might also contribute to the lower prediction skill of PDO.

This study provides valuable insights into the S2D prediction skills of ENSO and PDO. The results highlight the importance of both model initialization and the accurate representation of external radiative forcing, particularly for longer lead times. Multi-model ensemble averaging proves beneficial for ENSO prediction, but less so for PDO. Future research could explore advanced post-processing techniques and investigate the idealized impacts of Pacific and Atlantic decadal variability using additional CMIP6 experiments to improve understanding and prediction of S2D climate variability.

The study's reliance on a specific set of CMIP5 and CMIP6 models may limit the generalizability of the findings. The chosen metrics for assessing prediction skill might not fully capture all aspects of predictability. The analysis focuses primarily on large-scale indices (NINO3.4 and PDO), and regional variations in prediction skill may not be fully explored. The study does not explicitly investigate the relative contributions of specific processes to ENSO and PDO predictability.