Understanding Perturbed Parameter Ensemble Sensitivities Using A Contrastive Learning Approach

This repository contains the processed figure data and Python plotting scripts associated with the study, “Understanding Perturbed Parameter Ensemble Sensitivities Using A Contrastive Learning Approach.” The study applies explainable contrastive learning to two 100-member Community Atmosphere Model version 6 (CAM6) perturbed-parameter ensembles that use the same 34 parameter settings and differ in their warm-rain microphysics scheme: KK2000 and TAU-ML. Monthly fields of total liquid water path, shortwave and longwave cloud radiative effects, outgoing longwave radiation, and absorbed shortwave radiation are mapped into a shared representation space and compared with observational climatologies. Linear evaluation, principal-component analysis, cosine-distance analysis, and integrated gradients are used to diagnose differences among the ensembles and observations and relate regional sensitivities to physical parameters. The archive is organized by manuscript figure (Figures 2–12). Each figure directory contains the processed input data, plotting script, and rendered figure required to reproduce the corresponding analysis. The products cover classification across network layers; PCA projections of encoder and final-layer embeddings; monthly representation-space structure; seasonal cosine distances; parameter correlations; mean and seasonal integrated-gradient attribution maps; cloud and radiation differences; and regional attribution–parameter sensitivities. The full raw climate-model output and intermediate neural-network feature stores are excluded because of their large storage volume. The included processed inputs are sufficient to recreate Figures 2–12.