A Machine Learning Approach to Emulation and Biophysical Parameter Estimation for Land Models
Land models are essential tools for understanding and predicting terrestrial processes and climate-carbon feedbacks in the Earth system, but uncertainties in their future projections are poorly understood. Parametric uncertainty in land models has been traditionally explored through experimentation with different parameter values to test how variations impact resulting model predictions. Undirected model calibration can be limited by the ambiguous and subjective nature of the hand tuning process. In this work, we design and implement a machine learning approach to globally calibrate a subset of biophysical land model parameters to observations of carbon and water fluxes using version 5 of the Community Land Model.
Our study provides an example of a mechanistic framework for evaluating parameter uncertainty in a complex global land model. We further identify key processes and parameters that are important for accurate land modeling. By utilizing machine learning to build a fast emulator of the land model, we can more efficiently optimize parameter values with respect to observations and understand the different sources of uncertainty contributing to predictions in terrestrial processes.
We use parameter sensitivity simulations and a combination of objective metrics to determine a subset of important biophysical parameters for further analysis. Using a perturbed parameter ensemble, we then train a series of artificial feed-forward neural networks to emulate land model output given parameter values as input. The trained networks are then used to estimate global optimal parameter values with greater computational efficiency than achieved by hand tuning efforts and increased spatial scale relative to previous studies optimizing at a single site.