loading page

Using Gaussian process regression and stable isotopologues of water vapor to estimate shallow convective moistening in the southeastern Pacific marine stratocumulus region
  • Joseph Galewsky,
  • Camille Risi,
  • Hélène Brogniez
Joseph Galewsky
University of New Mexico

Corresponding Author:[email protected]

Author Profile
Camille Risi
Laboratoire de Météorologie Dynamique, France
Author Profile
Hélène Brogniez
Laboratoire atmosphères, observations spatiales
Author Profile


Convective mixing in the lower free troposphere (LFT) and its response to climate change are at the heart of low-cloud feedbacks in projections of future warming, but are challenging to diagnose from observations. The stable isotopic composition of water vapor in the LFT is a sensitive recorder of shallow convective moistening, and can potentially provide independent constraints on shallow convective processes. In-situ and remote sensing measurements from the southeast Pacific marine stratocumulus region and an isotope-enabled general circulation model (GCM) are used along with Gaussian process regression (GPR) to explore the utility of stable isotope measurements and simulations for improved estimates of shallow convective moistening tendencies in marine stratocumulus settings. We train the GPR algorithm on conventional and isotopic fields from a GCM (LMDZ5B) from the SE Pacific marine stratocumulus region and assess the algorithm on out-of-sample GCM output. The GPR trained on isotopic fields yields better estimates of shallow convective moistening tendencies than GPR trained only on conventional meteorological fields. Climate change is not well-captured if the GPR is trained only on the control climate, but performs much better if the training data include samples from both cool and warm climates, and is also reasonably well-captured if the GPR is only trained on the warm climate. The GPR algorithm is applied to isotopic and conventional measurements from the SE Pacific and yields realistic estimates of shallow convective moistening tendencies. Linking machine learning with isotopic simulations and measurements provides a unique and potentially useful framework for bridging GCMs and observations.