Constraining tropospheric stability trends using surface temperature patterns: A machine learning approach

The inconsistency between cloud feedbacks inferred from the historical period and from future warming scenarios can be attributed to their dependence on tropospheric stability, which itself is governed by evolving spatial patterns of surface warming. Changes in estimated inversion strength (EIS) with global warming modulate the strength of this so-called pattern effect, implying a critical need for observational constraints on how EIS responds to warming. However, changes in EIS with warming vary widely among different reanalysis datasets, limiting our ability to observationally constrain the pattern effect. In this study, we utilize machine learning to relate the spatial pattern of surface temperature warming to EIS changes between 1979 to 2022. The machine learning framework is trained by maps of surface warming from large ensembles of global climate simulations to predict changes in EIS per degree of global warming. The latter includes two components: 1) the forced EIS changes with warming (obtained from the ensemble mean) and 2) the unforced EIS changes with warming (obtained from deviations from the ensemble mean). The trained machine learning framework allows us to assess the credibility of reanalysis-derived EIS changes in light of observed surface temperature trends and additionally provides observational constraints on the relative importance of external forcing versus natural variability in causing recent EIS trends.