Understanding how land and atmosphere interact is critical for predicting climate change. Land surface models (LSMs) simulate how energy, water, and carbon move between the land and the air, but they often struggle with uncertainty and require a lot of computer power. To tackle these problems, we developed ClimaLand, a next-generation land model designed to run efficiently on graphics processing units (GPUs) and to combine physics-based modeling with data-driven learning.

ClimaLand is built to be flexible and fast. It allows scientists to include detailed real-world data where traditional models rely on guesswork. In this study, we focused on improving how the model simulates energy and water flows, such as how plants control water loss or how soil holds moisture. We used satellite-based climate data (ERA5) to adjust model parameters using a method called ensemble Kalman inversion, and we used machine learning to speed up the process.

We tested ClimaLand’s performance using an international benchmarking system and compared it to leading climate models. The results show improved accuracy with much faster calibration. ClimaLand paves the way for a new generation of climate models that are both more realistic and more scalable, which are key steps toward better climate predictions.