ClimaLand: Advancing Land Surface Modeling with Data-Driven Calibration and GPU Acceleration

Land surface models (LSMs) play a pivotal role in Earth System Models by simulating energy, water, and carbon fluxes between the land and the atmosphere. However, existing LSMs face challenges with computational efficiency and the calibration of uncertain parameterizations, particularly for key carbon and water fluxes. To address these limitations, we introduce ClimaLand, a GPU-native LSM designed to integrate machine learning (ML)  parameterizations and calibration frameworks with physical models. ClimaLand's modular architecture allows seamless incorporation of data-driven approaches for unresolved processes, such as subgrid-scale hydrology and canopy-atmosphere coupling, for faster iterations and hypothesis testing.

In this study, we focus on calibrating the latent heat flux, or evapotranspiration, a major source of uncertainty in land-atmosphere interactions. Using observational data from flux towers and remote sensing, we demonstrate how ClimaLand employs Ensemble Kalman Processes (EKP) to optimize parameterizations of stomatal conductance and soil moisture evaporation. Calibration approaches reduced bias during extreme events compared to traditional LSMs.

Benchmarking on GPUs highlights ClimaLand’s computational efficiency, enabling rapid uncertainty quantification and parameter ensemble testing. Results showcase the model’s capacity to improve physical realism and predictive accuracy, particularly for water and energy cycles critical to climate risk assessments.

ClimaLand marks a step forward in leveraging modern computational tools and ML to enhance the accuracy and scalability of LSMs. Future developments will extend to optimality-submodels and increased spatial resolution.