ArcticBench and SmartTransfer: Benchmarking and Enabling Continual Learning of Atmosphere Generative Foundation Model

Fu-Ming Guo, fengwei guo, Yingfang Fan

Keywords: Continual Learning, Atmosphere Generative Foundation Model, Image Superresolution, Domain Adaption, Earth Science

TL;DR: We introduce ArcticBench, a rigorous polar-specific evaluation grid, and SmartTransfer, a structure-aware weight transfer algorithm, solving brittle domain adaptation in atmospheric diffusion models.

Abstract:

Diffusion-based atmospheric generative foundation models offer a compelling route to uncertainty-aware super-resolution and downscaling,
but progress is constrained by two practical gaps: (i) the lack of a rigorous, reproducible benchmark tailored to polar conditions, and
(ii) brittle adaptation of pretrained models across domains, resolutions, and variable inventories.
We address both challenges with \textbf{ArcticBench} and \textbf{SmartTransfer}. ArcticBench is a curated benchmark built from the Copernicus Arctic Regional Reanalysis (CARRA), reprojected onto a hierarchical equal-area HEALPix grid, and paired
with versioned Arctic masks and Arctic-stratified metrics that probe regional skill, extreme-event fidelity, and spatial realism.
SmartTransfer is a structure-aware weight transfer algorithm for continual learning of conditional diffusion super-resolution models: it aligns resolution-dependent positional representations, selectively reuses domain-invariant positional channels in the encoder, copies the shared denoising backbone, and reinitialises irreducibly variable-specific heads. On ArcticBench, SmartTransfer improves both optimisation and test performance, reducing best validation loss by 42\% compared to training from scratch and achieving lower Macro-RMSE/MAE (10.97/4.03 vs 11.13/4.15). Crucially, gains concentrate in high-impact regimes, improving extreme-quantile RMSE by 22.6\% at $q{=}0.95$ and 41.6\% at $q{=}0.99$, with consistent per-variable benefits (e.g., $-27.1\%$ on skin temperature). Together, ArcticBench and SmartTransfer provide a reproducible polar evaluation standard and a practical mechanism for adapting atmospheric diffusion models under real-world configuration shifts.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 50