Skillful multiyear predictions of ocean acidification in the California Current System

Ocean acidification, driven by anthropogenic CO₂ uptake, lowers ocean pH and affects marine ecosystems. The CCS is productive and economically important but naturally prone to low pH due to wind-driven upwelling and Ekman suction that bring carbon- and nutrient-rich waters to the surface. This elevates biological productivity but also introduces corrosive conditions, exacerbated by anthropogenic CO₂, with impacts on calcifying organisms and fisheries. While seasonal forecasts of physical and some biogeochemical variables exist for CCS, there has been little work on multiannual to decadal prediction of ocean biogeochemistry. Such horizons are valuable for fisheries management (e.g., annual catch limits, closed areas, quotas). Persistence forecasts provide some short-lead skill but are limited by system decorrelation timescales. The study’s research question is whether initialized global Earth system model forecasts can skillfully predict interannual surface pH anomalies in the CCS beyond seasonal scales, outperforming persistence and externally forced, uninitialized simulations, and to identify the mechanisms that confer predictability.

Prior CCS prediction efforts emphasized seasonal forecasts of sea surface temperature and certain biogeochemical properties for ecosystem models, with emerging 1–2 year initialized predictions of surface chlorophyll. Decadal biogeochemical forecasting remains nascent. Persistence forecasts are commonly used baselines, providing skill at lead times near the system’s decorrelation times, but lack physically based evolution. Initialized decadal Earth system model ensembles, which couple atmosphere, ocean, cryosphere, land, and biogeochemistry under external forcing, offer potential to improve multiannual prediction by advancing realistic initial conditions. Previous work shows predictability in marine productivity and carbon uptake on multiyear scales, motivating exploration of ocean acidification predictability in the CCS.

- Prediction system: Community Earth System Model Decadal Prediction Large Ensemble (CESM-DPLE; CESM v1.1 with BEC ocean biogeochemistry). Ocean model nominal 1°×1° with 60 vertical levels; coupled atmosphere (CAM5, ~1°, 30 levels), land, sea ice, and ocean components. - Initialization and ensembles: Full-field initialization annually on November 1 for 1954–2017 (64 start dates). Forty-member ensembles created by tiny Gaussian perturbations to atmospheric temperature; other components evolve spread from atmospheric perturbations. Each member integrated 122 months (~10 years). Total ~26,000 simulation years. - Forcing and controls: Historical radiative forcing (including volcanic) through 2005; scenario-based projected forcing thereafter. Uninitialized comparison from CESM Large Ensemble (CESM-LE) with identical code/forcing, initialized once. - Reconstruction (verification for potential predictability): Forced ocean–sea ice reconstruction (1948–2017) with identical ocean/ice resolution; CORE interannual fluxes, with tropical wind corrections using NOAA 20CRv2 and JRA-55-do to remove spurious SST trend. No direct ocean/ice data assimilation; skill arises from atmospheric reanalysis forcing. Reconstruction shows realistic CCS surface pH mean state, seasonality, and variability versus observations. - Observational product (skill verification): JMA Ocean CO₂ Map (1°×1°, 1990–2017), providing monthly surface pH derived diagnostically from empirical alkalinity (from SSS and SSH by five-region regressions) and pCO₂ (from regional multivariate regressions on SST, SSS, Chl-a across 44 regions) fed into a carbonate system solver. - Domain and preprocessing: Analyses over the California Current Large Marine Ecosystem (area-weighted). Focus on surface pH anomalies; remove a second-order polynomial fit (to filter long-term acidification trend) and seasonal cycle as appropriate. Annual forecasts defined as January–December following Nov 1 initialization (lead year 1 covers months 3–14). Drift adjustment applied to initialized forecasts by subtracting lead-dependent model climatology computed across members/start dates (hindcasts verifying 1964–2014). Anomalies for CESM-LE/reconstruction over 1964–2014; JMA over 1990–2005 for verification window. - Metrics and statistics: Predictability/potential predictability quantified by anomaly correlation coefficient (ACC) between forecasts and reconstruction; predictive skill by ACC vs JMA observations (1990–2005). Accuracy quantified by normalized mean absolute error (NMAE), MAE normalized by observed/reconstruction interannual standard deviation. Significance of ACC via t-test using effective sample size to account for autocorrelation; differences in ACC (initialized vs persistence/uninitialized) via z-test (95% confidence). Persistence forecast assumes anomalies at initialization persist to all leads. Uninitialized forecast uses CESM-LE ensemble mean anomalies. - Model evaluation: Reconstruction vs JMA (1990–2005): spatial climatologies and seasonal cycles consistent; linear pH trends: −0.026 (recon) vs −0.029 (obs); interannual std ~0.003 pH; monthly anomaly correlation r=0.72; decorrelation timescale ~4 months; slight acidic bias (H+ bias 2.9–4.2%). CESM reproduces alongshore winds and large-scale air–sea CO₂ flux patterns; nearshore upwelling-scale CO₂ outgassing not resolved at 1°. - Mechanism diagnostics: Linear decomposition of pH variability and predictability into SST, SSS, salinity-normalized DIC (sDIC), and salinity-normalized alkalinity (sALK), scaled to pH units using sensitivities from CO2SYS. DIC budget for upper 150 m over CCS, decomposed into vertical/lateral advection, vertical/lateral mixing, biology, and gas exchange; correlations among terms assessed to identify controls of DIC tendency.

- Skill relative to reconstruction (potential predictability): Initialized forecasts show robust ACC skill for annual surface pH anomalies across the CCS, exceeding persistence and uninitialized forecasts out to 5-year leads. Area-weighted ACC at lead year 1 is 0.72 (explaining >50% variance); initialized forecast remains statistically significant over persistence and uninitialized through 5-year leads. Persistence yields positive ACC only at lead 1 in parts of the CCS, becoming weak/negative by years 2–5. - Spatial patterns (reconstruction verification): Initialized ACCs remain significant in central and southern CCS through year 5; persistence fails (negative ACC) over most of CCS beyond year 1. AACC (initialized minus persistence ACC) is positive across nearly all grid cells through year 5 (except a few coastal Pacific Northwest cells at year 3). - Accuracy: NMAE of initialized forecasts is consistently below that of persistence and uninitialized forecasts and within the interannual variability of reconstruction across all 10 lead years. - Skill vs observations (JMA, 1990–2005): Positive initialized ACC across most CCS at lead 1; statistically significant improvements over observational persistence from Cape Mendocino to Baja California through lead years 1–4. Observational persistence is marginally useful south of Cape Mendocino at lead 1 but generally negative ACC by years 2–5. Area-weighted AACC ranges ~0.04–0.43 across leads 1–5. Skill wanes earlier in the southernmost CCS (by year 2) and Pacific Northwest (by year 3). Initialized NMAE is lower than observational persistence across most of the CCS for leads 1–5, typically below unity (within observed interannual variability). - Mechanisms: Predictability of surface pH is dominated by predictability of salinity-normalized dissolved inorganic carbon (sDIC), contributing the largest scaled pH predictability across 10 lead years; combined SST and sALK predictability is comparable to sDIC over the first 5 years; SSS contributes negligibly. sDIC predictability arises largely from persistence of initialized anomalies and is enhanced by initialization. Upper-150 m DIC budget shows variability primarily set by vertical and lateral advection (correlation between advective flux and total DIC tendency r≈0.9). Source waters (subarctic via California Current; eastern tropical Pacific via California Undercurrent) carry interannual–decadal biogeochemical anomalies into the CCS, underpinning multiyear predictability. In the reconstruction, DIC and SST components correlate with PDO (r=0.66 and 0.73) and ENSO (r=0.52 and 0.64), while net pH anomalies (a small residual of many terms) show near-zero correlations with these modes. - Additional quantitative evaluation: Reconstruction vs JMA monthly pH anomalies correlate at r=0.72 (1990–2005). Linear pH trends over 1990–2005: −0.026 (recon) vs −0.029 (obs). Interannual standard deviation of pH is 0.003, 1.5–2× the annual trend magnitude; decorrelation timescale ~4 months.

The study demonstrates that an initialized Earth system model can skillfully predict interannual CCS surface pH anomalies beyond seasonal horizons, with practical skill at one year and potential skill up to five years. Initialization of subsurface and basin-scale DIC anomalies—transported into the CCS by meridional and vertical advection—provides the primary source of predictability. By outperforming persistence and uninitialized, externally forced simulations, the results confirm that realistic initial ocean biogeochemical states are crucial for multiyear ocean acidification predictions. These forecasts can inform expectations of temporary accelerations or slowdowns in the ongoing acidification trend, and the initialized global ESM outputs could serve as boundary conditions to enhance skill and extend lead times in high-resolution regional models. Such downscaled systems could better resolve coastal processes relevant to ecosystems and fisheries management, and other carbonate system metrics (e.g., aragonite saturation state) are likely predictable due to their dependence on dissolved CO₂ variability.

Initialized CESM-DPLE forecasts provide skillful predictions of CCS surface pH anomalies at least one year ahead, with potential predictability up to five years, surpassing persistence and uninitialized approaches. Predictability is primarily enabled by initialized sDIC anomalies and their advective transport into the CCS. These findings establish the feasibility of multiyear ocean acidification prediction in a complex, vulnerable, and economically important region, and point to the utility of global initialized forecasts as boundary conditions for regional biogeochemical prediction and for informing management about interannual fluctuations around the acidification trend. Future work should incorporate higher-resolution dynamical downscaling to resolve nearshore upwelling, enhance observational and reanalysis constraints on DIC and transport fields, extend verification with improved observational products, and explore predictability of related carbonate system variables.

- Model resolution: ~1° (~100 km) grid cannot explicitly resolve nearshore coastal upwelling (within ~30 km of coast) and uses parameterizations for eddy-induced tracer fluxes; this suppresses pH variability and limits direct applicability to coastal fisheries management without downscaling. - Target anomaly magnitude: Annual surface pH anomalies are small (<0.01 units), though associated with larger variations in related metrics (e.g., aragonite saturation ~0.1 units); nevertheless, forecast errors generally remain within historical variability. - Temporal resolution: Analysis emphasizes annual means because monthly predictability is lower; monthly skill exists mainly into the first upwelling season(s) post-initialization. - Detrending: Analyses remove the long-term acidification trend to isolate interannual variability; results are similar with trends retained but the focus here is variability around the trend. - Observational verification window: JMA product spans 1990–2017, but reconstruction lost observed atmospheric CO₂ variability after 2005 (scenario projection), reducing correspondence with observations; thus, skill was assessed over 1990–2005, limiting degrees of freedom for significance testing. - Reconstruction forcing and data: No direct ocean/sea-ice assimilation; skill derives from atmospheric reanalysis forcing and model dynamics. Uncertainties in empirical JMA pH derivation (no formal pH uncertainty provided) also constrain evaluation. - Operational needs: Skill relies on initializing subsurface DIC and transport; operational forecasting would benefit from enhanced observations/reanalyses of DIC, alkalinity, and circulation to improve initial states.