Neural emulator based on physical fields for accelerating the simulation of surface chlorophyll in an Earth System Model

Simulating the ocean biogeochemical module (BGC-enabled) in the Community Earth System Model (CESM) is computationally expensive, often requiring significantly more resources than the physical climate component. In this study, we propose an alternative approach to generate biogeochemical data using a neural network emulator, BGC-UNet, which predicts ocean surface chlorophyll concentrations based on physical fields from CESM, such as solar short-wave heat flux (SHF-QSW), potential temperature (TEMP), and zonal and meridional velocity (UVEL, VVEL). BGC-UNet is designed as a UNet-like architecture and employs a patch-based methodology with dilated sampling to efficiently reconstruct biogeochemical data from physical inputs. This framework potentially enables high-resolution chlorophyll predictions without running full BGC-enabled simulations. Our evaluation demonstrates that BGC-UNet’s outputs closely align with CESM’s simulated surface chlorophyll, supported by both quantitative metrics and visual analysis. Additionally, the emulator achieves a simulation speed approximately 248 times faster than traditional BGC-enabled CESM simulations. Although the current focus is on surface chlorophyll, the model shows potential for future extension to other biogeochemical variables. By leveraging only 40 years of simulated data for training, BGC-UNet replicates the trends observed in CESM, making it a promising tool for accelerating Earth system modeling.