Cause of decreased water‑vapor tendency skill above the tropopause in the constrained U‑Net

Determine the underlying causes of the reduced offline coefficient of determination (R^2) for predicted water vapor tendencies above the tropopause when using the U‑Net‑expanded‑constrained architecture that enforces temperature‑based liquid–ice partitioning and total cloud prognostics, trained on the ClimSim dataset and evaluated against E3SM‑MMF cloud‑resolving model outputs.

Background

The paper introduces a U‑Net architecture with microphysical constraints—specifically predicting total cloud condensate and enforcing a temperature‑based liquid–ice partition—to emulate the cloud‑resolving model (CRM) within E3SM‑MMF. This approach improves the prediction of cloud condensate tendencies in cold upper‑tropospheric regions where liquid clouds should be absent.

However, the authors report that, despite these improvements, the constrained U‑Net exhibits a decrease in R^2 for water vapor tendencies above the tropopause. While they speculate that using only total cloud mixing ratio inputs might omit relevant information for predicting water vapor tendencies at those levels, they explicitly state that the cause of the degradation remains unclear.