TorchClim v1.0: a deep-learning plugin for climate model physics

• Published in: Geoscientific Model Development 2024-07, Vol.17 (14), p.5459-5475
• Main Authors: Fuchs, David, Sherwood, Steven C, Prasad, Abhnil, Trapeznikov, Kirill, Gimlett, Jim
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Atmospheric models; Climate; Climate models; Conservation laws; Deep learning; General circulation models; Hybrid systems; Machine learning; Neural networks; Parameterization; Physics; Precipitation; Radiation
• ISSN: 1991-9603; 1991-959X; 1991-962X; 1991-9603; 1991-962X
• DOI: 10.5194/gmd-17-5459-2024

Climate models are hindered by the need to conceptualize and then parameterize complex physical processes that are not explicitly numerically resolved and for which no rigorous theory exists. Machine learning and artificial intelligence methods (ML and AI) offer a promising paradigm that can augment or replace the traditional parameterized approach with models trained on empirical process data. We offer a flexible and efficient plugin, TorchClim, that facilitates the insertion of ML and AI physics surrogates into the climate model to create hybrid models. A reference implementation is presented for the Community Earth System Model (CESM), where moist physics and radiation parameterizations of the Community Atmospheric Model (CAM) are replaced with such a surrogate. We present a set of best-practice principles for doing this with minimal changes to the general circulation model (GCM), exposing the surrogate model as any other parameterization module, and discuss how to accommodate the requirements of physics surrogates such as the need to avoid unphysical values and supply information needed by other GCM components. We show that a deep-neural-network surrogate trained on data from CAM itself can produce a model that reproduces the climate and variability in the original model, although with some biases. The efficiency and flexibility of this approach open up new possibilities for using physics surrogates trained on offline data to improve climate model performance, better understand model physical processes, and flexibly incorporate new processes into climate models.