Earth System Modeling with Coupled Convolutional Neural Networks

Abstract

This dissertation advances the application of deep learning (DL) to Earth system modeling. Chapter 2 introduces DLESyM (Deep Learning Earth SYstem Model), a fully data-driven coupled atmosphere-ocean model. DLESyM pairs a deep learning weather prediction module (DLWP) with a deep learning ocean module (DLOM), linked through an asynchronous coupling scheme that mirrors the design of numerical Earth system models. DLESyM uses a parsimonious set of prognostic variables — nine atmospheric fields and sea surface temperature — yet sustains stable, physically coherent rollouts over millennial timescales without appreciable drift. Free-running simulations reproduce a wide range of observedclimate phenomena, including extratropical atmospheric blocking, annular modes, and El Ni˜no–Southern Oscillation. DLESyM performs comparably to or better than several leading CMIP6 models in reproducing these phenomena when compared to ERA5 observations over the historical period. Chapter 3 documents DLESyM’s contribution to the inaugural Artificial Intelligence Model Intercomparison Project (AIMIP), a coordinated community effort to standardize the evaluation of AI-based climate models. DLESyM contributed fifteen simulations spanning historical, +2K, and +4K SST perturbation experiments. The chapter highlights DLESyM’s behavior in these simulations and its performance relative to other participating models and a reference numerical model. Chapter 4 extends the modeling framework to the land surface, introducing a data driven land and terrestrial model (DLTM) that simulates vegetation dynamics and land surface response to atmospheric forcing. DLTM is trained on ERA5-Land and satellite derived NDVI, and is shown to reproduce observed land surface climatology. We include analysis of DLTM’s response to extreme heat events in the atmosphere — phenomena for which land-surface memory is known to be a critical but often poorly represented source of predictability. In Chapter 5 we present preliminary results assessing the effect of land coupling for a simple deep learning weather prediction model. Taken together, these contributions establish that component-wise, coupled deep learning modeling is a viable paradigm for simulating the Earth system and capable of capturing emergent climate variability with a fraction of the computational cost of numerical models.