B. B. Cael, Friedrich Burger, Gregory Britten, Stephanie Henson, Thomas Frolicher

Marine heatwaves impact ocean ecosystems and are expected to become more frequent and intense with continued global warming. The ability of Earth system models to reproduce the statistical characteristics of extreme ocean temperatures has not yet been tested quantitatively, making the reliability of their future projections of marine heatwaves uncertain. We demonstrate that annual maxima of detrended anomalies in daily-mean sea surface temperatures over the last 39 years of global satellite observations are described excellently by the Generalised Extreme Value (GEV) distribution, as predicted from extreme value theory. GEV parameters' spatial patterns conform to physical expectations, further supporting its use for model-observation comparison. Historical realisations of 14 CMIP6 Earth system models reproduce the GEV and spatial patterns in the underlying parameters. We can then use these models with confidence to project future changes in maximum ocean temperatures, which we show will become warmer (by 1.08$\pm$0.18$^\circ$C on average under 2$^\circ$ warming and 2.06$\pm$0.19$^\circ$C on average under 3.2$^\circ$C warming) and tend to increase more than global mean sea surface temperature (0.92$\pm$0.18$^\circ$C and 1.77$\pm$0.14$^\circ$C respectively). Our study provides an effective means to quantify extreme ocean temperatures, as well as confidence in the predictions of future marine heatwaves from CMIP6 models.