Interpreting and Stabilizing Machine-Learning Parametrizations of Convection

• Published in: Journal of the atmospheric sciences Vol. 77; no. 12; pp. 4357 - 4375
• Main Authors: Brenowitz, Noah D., Beucler, Tom, Pritchard, Michael, Bretherton, Christopher S.
• Format: Journal Article
• Language: English
• Published: Boston American Meteorological Society 01.12.2020
• Subjects: Atmospheric convection; Atmospheric models; Climate; Climate models; Cloud resolving models; Clouds; Computational fluid dynamics; Conditional instability; Control stability; Drying; Dynamic stability; ENVIRONMENTAL SCIENCES; Fluid dynamics; Gravity waves; Growth rate; Hydrodynamics; Instability; Learning algorithms; Learning behaviour; Machine learning; MATHEMATICS AND COMPUTING; Meteorology & Atmospheric Sciences; Moist convection; Neural networks; Parameterization; Phase velocity; Physics; Response functions; Spectral analysis; Spectrum analysis; Stability analysis; Stabilizing; Wave dynamics; Wavelengths

Neural networks are a promising technique for parameterizing subgrid-scale physics (e.g., moist atmospheric convection) in coarse-resolution climate models, but their lack of interpretability and reliability prevents widespread adoption. For instance, it is not fully understood why neural network pa...