2026/06/02 by David Fuchs, Steven C. Sherwood, Abhnil Prasad +1 · 1 voice
Earth and Planetary Sciences · Environmental Science · #Climate variability and models #Meteorological Phenomena and Simulations #Tropical and Extratropical Cyclones Research
paper · pdf · doi:10.1175/aies-d-25-0115.1
openalex publication_date 2026/06/02 · openalex created_date 2026/06/03 · openalex updated_date 2026/07/31
Abstract Numerous studies have explored the use of neural networks trained to serve as surrogates for predicting subgrid physical processes in atmospheric models solving physical equations. Here, we explore three extensions to a residual network (ResNet) regression backbone architecture, adding depth and dense connections and combining both into a mesh of shallow and deep lateral pathways in place of dense connections, referred to as mesh directed acyclic graph (MeshDAG). The networks were used to emulate moist processes in the standard configuration of the Community Atmosphere Model, specifically to predict convective and boundary layer tendencies of temperature, humidity, and condensed liquid and ice, as well as convective precipitation. In spite of uniform sampling of training data and a relatively simple loss function, all architectures mastered tropical and liquid water processes during earlier training epochs. Except for MeshDAG, architectures struggled in later training epochs in learning extratropical and mixed state water tendencies, exhibiting overfitting and numerical instabilities instead. A MeshDAG hybrid climate model successfully ran for 25 years as an atmosphere–land model and reproduced well the behavior of the emulated model. A hybrid atmosphere–ocean model also ran stably for 12 years, albeit with small climate drifts due to biases in simulated mid- and high-latitude cloud cover and hence top-of-the-atmosphere radiation balance, induced by remaining biases in surrogate-predicted condensed water tendencies. This work suggests that alternative architectures with lateral connections may prove more robust for emulating physical processes than simply going to deeper networks and suggests further promise for hybrid modeling of climate, including prognostic condensed-water fields. Significance Statement Hybrid climate models, where physical equations are augmented by artificial intelligence surrogates of hard-to-model climate system components, hold promise to improve global climate simulation. However, they are beset by a number of challenges, including the performance of the surrogates, which are usually some type of neural network. We show that adding lateral (both shallow and deeper) pathways to such a network can significantly improve the skill and runtime stability, and we present the first surrogate to model the full set of outputs of a traditional numerical moist convection and run stably in a climate model with active ocean, atmosphere, and land.