Perturbed parameter ensembles (PPEs) expose how physical parameters influence climate simulations, yet interpreting parameter sensitivities across multivariate, spatially structured outputs remains difficult, especially when calibrating models to observations. We develop an explainable contrastive learning model that maps five monthly cloud and radiation fields into a shared representation space. The model is trained on two 100‑member Community Atmosphere Model version 6 (CAM6) PPEs, covering 34 parameters and differing only in the warm‑rain microphysics scheme: the default bulk scheme KK2000 and TAU‑ML, a neural‑network emulator of a bin microphysics scheme. The learned representations separate the two PPEs with over 94% linear classification accuracy while preserving seasonal variability and ensemble spread caused by parameter perturbations. In the shared space, satellite observation representations lie on the same low‑dimensional manifold as the PPEs but are most displaced during boreal spring and autumn. The TAU‑ML PPE consistently shows a smaller distance to observations than KK2000. Integrated Gradients attributions highlight contributions from subtropical low‑cloud regions, Northern and Southern Hemisphere storm tracks, and tropical convection zones to the differences between PPEs and observations. These regional attributions correlate strongest with parameters governing cloud microphysics, boundary‑layer turbulence, and deep convection. The findings demonstrate that explainable representations of climate fields can attribute model differences to specific variables, regions, seasons, and physical parameters.
Review