Future Warming Shortens the Memory of the Pacific Decadal Oscillation and Erodes a Source of Near-Term Predictability

The decadal predictability of the Pacific Decadal Oscillation (PDO) rests in part on the memory of North Pacific sea surface temperature (SST) anomalies. Comparing the historical experiment and four Shared Socioeconomic Pathways (SSPs) of eleven CMIP6 models with control-run segments of equal length, we find that the year-to-year persistence of the PDO declines under every SSP, most clearly under SSP3-7.0. Under SSP3-7.0 the lag-1 autocorrelation of annual means falls from 0.49 to 0.35, in all eleven models and by 4.2 times the spread expected from internal variability, while the historical period shows no change. Two 50-member large ensembles, of CESM2 and MIROC6, reproduce the decline. PDO variance moves from decadal to interannual periods at nearly unchanged total variance, and although ENSO variance grows by 30\% to 60\%, most of the persistence loss survives ENSO removal. In the PDO core region the winter mixed layer shoals in ten of eleven models and in every ensemble member, explaining about half of the loss. With the change in turbulent air-sea damping, which increases in most models by a model-dependent amount, shoaling accounts for most of it in a stochastic-model budget, in the multi-model mean (observed over predicted change 1.15, 95\% interval 0.62 to 1.89) and in the ensemble. Persistence forecasts and a linear inverse model of the North Pacific lose skill, while tropical Pacific information offsets part of the loss at leads up to a year. Analogue-based predictability limits from single 86-year runs miss the change.