12,273 Publications · 71,443 Citations

Question: after 2012, did surface ocean carbonate chemistry at the two longest open-ocean time series (HOT, North Pacific, 1988-2024, 346 cruises; BATS, Sargasso Sea, 1991-2025, 378 cruises) follow frozen trend models, an atmospheric-tracking model, or tracking plus the thermodynamic temperature effect? Approach: pre-registered forward test. We fitted models on data to 2012 (check: to 2007), recalculated the carbonate system from DIC and TA with PyCO2SYS, and scored test-period bias and RMSE with year-block bootstrap intervals. Results: seawater pCO2 rose 2.74 (HOT) and 3.14 (BATS) uatm/yr in 2013-2025, against 2.42 and 2.51 in the atmosphere. The extra sea-air trend depends on the split year. Atmospheric tracking cut aragonite saturation forecast RMSE by 19-31% at both stations and splits. For pCO2 it did not beat a linear trend, and for pH the result was mixed. Pre-registered H3 failed: adding the observed temperature anomaly (test years about 0.5 C warmer) made all pCO2, pH and saturation forecasts worse (HOT pCO2 RMSE 7.5 to 14.2 uatm). Exploratory: annual-mean temperature anomalies passed 0.00 [-0.23, 0.23] (BATS) and -0.14 [-0.33, 0.12] (HOT) of the thermodynamic effect to pCO2. Within-year deviations passed 0.58 and 0.37. At both stations, TA rose above the pre-2013 TA-salinity line until about 2013-2016 and then fell. This adds about 0.5-0.7 uatm/yr to calculated pCO2, the same size as the extra post-2012 pCO2 acceleration. At HOT, measured spectrophotometric pH fell at -0.00231/yr, as atmospheric tracking predicts (-0.00237), while pH calculated from DIC and TA fell at -0.00272/yr (difference trend -0.00041/yr [-0.00064, -0.00019]). Limitations: two subtropical stations; 12-13 test years with large interannual noise; pass-through is correlational; we cannot tell if the TA reversal is oceanographic or analytical (shared reference-material calibration). External data are not bundled; scripts download them.

Question: did global mean sea level (GMSL) after 2017 follow the quadratic extrapolation of Nerem et al. (2018), and was a quadratic a better 8-year forecast than a linear trend? Approach: we fitted forecasts to the exact University of Colorado (CU) release 2018_rel1 (1993-2017; raw quadratic 0.097 mm/yr^2, the paper value) and compared them with CU 2026_rel2 (to 2026.46) and NOAA LSA (to 2025.12), relative to the 2013-2017 mean. Results: the published climate-driven path (2.9 mm/yr at 2005, 0.084 mm/yr^2) had mean error -1.2 mm (CU) and +0.1 mm (NOAA). The raw quadratic forecast too much rise (last-12-month error -9.7 mm; 3.3 mm of this comes from data reprocessing, which lowered the 1993-2017 acceleration to 0.083). A linear 1993-2017 fit forecast too little (+5.7 to +9.0 mm). The 2018-2026 rate was 3.87 mm/yr (CU) against 4.35-4.75 mm/yr from accelerating forecasts. AR(12) and block-bootstrap simulations of a quadratic law put these errors inside interannual noise (one-sided p 0.12-0.24), and the full-record acceleration fell from 0.088 (to 2020) to 0.074 mm/yr^2 (to 2026), also inside noise. In rolling-origin hindcasts (2003-2018) the quadratic beat the linear fit at only 11 of 31 origins, but at all origins from 2014.5. A diagnostic with GRACE/GRACE-FO mascons shows a slower ocean-mass gain in 2018.5-2026.5 (1.30 +/- 0.25 vs 2.13 +/- 0.13 mm/yr) and faster 0-700 m steric rise. Limitations: 8.5-year verification, simple Nino3.4 adjustment (not the CSEOF method), reprocessed releases, and no closed budget (no atmospheric-mass correction).

Question: does the height of high water above the annual mean sea level change, and is the cause the astronomical tide or the non-tidal part? Approach: we used UHSLC/JASL Research Quality hourly sea level at 173 stations (6,528 station-years, 1980-2021). For each station-year we fitted 31 tidal constituents by least squares. The primary metric was the 99th percentile of hourly sea level above the annual mean (P99). Trend models included 18.61-year and 4.43-year terms. A joint-year block-permutation null kept the correlation between stations. Results: the network median P99 trend was +0.17 mm per decade (95% interval -0.62 to +1.16). The median absolute P99 trend was 0.11 times the median absolute relative mean sea-level trend. But 35 stations had |t| > 1.96, against a null 95th percentile of 22 (p = 0.004). Models fitted on 1980-2009 with a linear trend had larger 2010-2021 errors than models without a trend (median MAE ratio 1.062, interval 1.033-1.089; trend better at 29% of stations). Tide trends explained more of the cross-station variance than non-tidal trends (R2 0.59 versus 0.32), but five large-tide stations caused this difference. Exploratory: on the US Atlantic and Gulf coasts (28 stations) P99 rose by a median 5.9 mm per decade (permutation p = 0.007), mainly in the 30-day-mean non-tidal part. This trend also did not extrapolate. Elsewhere the fast non-tidal part and skew surge decreased (p = 0.003 and 0.016), and instrument changes can explain this. Limitations: uneven network with few European stations, data end in 2021, one-year tide fits, post hoc regions.