Kim et al. (2026) Deep learning in reduced-order space accelerates high-resolution multi-model climate downscaling

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2026
• Date: 2026-09-24
• Authors: Donghyuk Kim, Hajoon Song, Yujin Kim, Yeji Choi, Junghee Yun
• DOI: 10.1038/s41612-026-01545-y

Research Groups

• Department of Atmospheric Sciences, Yonsei University (Donghyuk Kim, Hajoon Song)
• Climate Intelligence Lab, DI Lab Inc. (Yujin Kim)
• Regional climate projections using dynamical models are computationally intensive, and generating high-resolution multi-model ensembles can be prohibitively expensive.

Short Summary

This study presents a lightweight deep-learning downscaling framework that generates high-resolution sea surface temperature (SST) projections within minutes on standard CPU-based systems. The framework operates in a reduced-dimensionality space by learning relationships between principal component (PC) time series derived from coarse-resolution climate model output and high-resolution dynamically downscaled SST.

Objective

• To develop a deep-learning downscaling framework that can generate high-resolution multi-model ensemble projections of sea surface temperature (SST) at regional scales.
• To evaluate the performance of the proposed framework using a set of 15 CMIP6 models and compare it with traditional dynamical downscaling methods.

Study Configuration

• Spatial Scale: Regional scale (northwestern North Pacific)
• Temporal Scale: 1995–2100 period

Methodology and Data

• Models used: Common Empirical Orthogonal Function (cEOF) analysis, Transformer model
• Data sources: CMIP6 datasets, ScenarioMIP scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5)

Main Results

• The proposed framework can generate high-resolution SST projections with a correlation coefficient exceeding 0.9 for the leading PC modes.
• The framework produces reliable downscaled SST that is comparable to the dynamically downscaled SST using the ocean model in a few minutes.
• The CEDL model effectively captures and translates coarse-resolution atmospheric forcing into high-resolution SST fields.

Contributions

• This study presents an efficient and reliable approach for high-resolution climate downscaling and uncertainty quantification.
• The proposed framework can generate large-scale prediction corresponding to each CMIP6 model, utilizing only the EOF patterns of the source RCM.
• The CEDL model is highly extensible to other ensemble-related research applications.

Funding

• This work was supported by the National Research Foundation of Korea (NRF) through the Basic Science Research Program (2019R1A2C3006924).
• This study was also funded by the Korean Ministry of Education and Human Resources Development through the Brain Korea 21 Plus Program.

Citation

@article{Kim2026Deep,
  author = {Kim, Donghyuk and Song, Hajoon and Kim, Yujin and Choi, Yeji and Yun, Junghee},
  title = {Deep learning in reduced-order space accelerates high-resolution multi-model climate downscaling},
  journal = {npj Climate and Atmospheric Science},
  year = {2026},
  doi = {10.1038/s41612-026-01545-y},
  url = {https://doi.org/10.1038/s41612-026-01545-y}
}