An Analysis of Training Strategies for Adapting Deep Learning-Based Emulators to New Geographic Regions

Global climate models, such as those available from the Coupled Model Intercomparison Project Phase 6 (CMIP6) project datasets, are often generated at a low resolution due to the large spatial and temporal scales of these models. This low resolution limits their application to areas where localized climate predictions are needed. One common method to increase this resolution is dynamical downscaling, where low-resolution climate forecasts are downscaled via regional numerical weather forecasting models, such as the Weather Research and Forecasting (WRF) model, to produce high-resolution results. While dynamical downscaling is an effective and physically reliable way to increase resolution, models like the WRF are computationally expensive and require long run times, limiting the ability to perform dynamical downscaling on large datasets (e.g., CMIP6).

In order to more efficiently downscale large datasets, we emulate the behavior of WRF using Generative Adversarial Network (GAN)-based deep neural network architectures trained on dynamically downscaled forecasts sampled from CMIP6. Generally, emulators trained in this fashion should only be applied within domains seen during training, which limits their utility outside the training domain. In this work, we explore strategies for training emulators that both more efficiently adapt to new domains and improve performance on unseen domains. In particular, continual and multitask learning strategies are applied to train and adapt an emulator on a known set of geographic regions. Both strategies demonstrate improved performance when multiple training domains are utilized. In particular, continual learning demonstrates the value of adapting to new domains and multitask learning improves generalization on domains unseen during training.