Transfer learning for estimating dynamic precipitation across different climate models

Variations in the global water cycle are among the most impact-relevant consequences of climate change to human society. However, large uncertainties exist in regional precipitation trends due to the presence of internal variability, limited coverage of observational data, and model deficiencies. Dynamical adjustment methods have been developed to quantify the role of circulation-induced variability in relevant climate variables such as precipitation and temperature. By subtracting the latter, the thermodynamic signals are expected to remain in the residuals, therefore increasing the signal-to-noise (S/N) ratio of trend estimates. However, dynamical adjustment for high-resolution spatial and temporal data remains challenging, especially over heterogeneous terrain such as the Alps. In this study, we build upon the Latent Linear Adjustment Autoencoder (LLAAE), a new statistical model introduced by Heinze-Deml et al., (2021, Geoscientific Model Development 14(8), pp. 4977-4999, doi:10.5194/gmd-2020-275.) that combines dynamical adjustment techniques with elements of deep learning. By combining a linear model with a nonlinear deep neural network, the method has shown great promise in extracting the forced thermodynamic components of high-resolution winter precipitation over Europe. So far, this model has only been trained and tested on different members of the Canadian Regional Climate Model (CanESM2). Here, we illustrate and evaluate the LLAAE in the context of transfer learning. The LLAAE, trained on 9 members of the CanESM2, is evaluated on 10 different high-resolution (EUR-11) Regional Climate Models of the EURO-CORDEX ensemble. We show, despite large climatological model differences, that the LLAAE is capable of reconstructing up to 50 % of the daily precipitation variance across the spatial domain of the CORDEX-data. Furthermore, trend estimates can be improved when retraining the linear model with relatively few additional data points from the target distribution: Retraining the model with 4000 individual days further increases the explained variance by about 30%, improving the S/N ratio of long-term and spatial trend estimates. This could be particularly useful for future real-world applications such as the detection and attribution of anthropogenic forced regional changes in precipitation and the contribution of the circulation-induced variability.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.