Multivariate deep-learning based regional climate model emulators and the impact of temporal awareness
Deep learning (DL) models have emerged as efficient approaches for producing high-resolution climate projections complementary to computationally expensive Regional Climate Models (RCMs). Many applications of high-resolution climate data, such as driving impact models or the characterization of extreme events, require multiple variables and an accurate representation of the relationships between them. Given that most DL-based RCM emulators are single-variate, this study investigates whether training multiple variables within the same model (i.e., multivariate emulators) can improve inter-variable coherence and emulator performance. This paper compares four multivariate RCM emulators and a single-variate baseline in their ability to reproduce daily 2D maps of average, maximum, and minimum surface temperature. All models rely on a U-Net-based architecture. Two models introduce temporality, one of which incorporates a ConvLSTM architecture, by using multiple days of input data. These models are evaluated within a perfect-model framework using a range of climate and numerical metrics. This study introduces a statistical significance testing framework to ensure that differences in model performance are not attributable to stochasticity in DL training. The results show that the temporal and multivariate emulators, especially the LSTM model, consistently outperform the other models across multiple metrics, while the non-temporal multivariate models perform similarly to the single-variate baseline.
Authors
- Christina Carty (ORCID: https://orcid.org/0009-0009-2089-6215)
- Julien Boé (ORCID: https://orcid.org/0000-0003-2965-4721)
- Antoine Doury (ORCID: https://orcid.org/0000-0002-4201-4414)
Institutions
- Centre National de la Recherche Scientifique (FR)
- Centre National de Recherches Météorologiques (FR)
- Centre Européen de Recherche et de Formation Avancée en Calcul Scientifique (FR)
- Météo-France (FR)
- Université de Toulouse (FR)
Publication Details
- Journal
- Artificial Intelligence for the Earth Systems
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1175/aies-d-26-0008.1
- Primary Topic
- Climate variability and models
- Type
- article
- Field-Weighted Citation Impact
- 0.00