Bayesian Calibration with Neural Network-Based Emulation of a Land Model

Abstract:

Land models simulate societally-relevant processes such as ecosystem dynamics, terrestrial hydrology, and agriculture with a high degree of complexity. Adequately predicting terrestrial processes such as climate-carbon feedbacks relies on assessing confidence in these models and their predictive capabilities while minimizing sources of model error. Parametric uncertainty in land models has been traditionally explored through experimentation with different parameter values to test how variations impact resulting model predictions. Hand tuning parameter values can be computationally inefficient, requiring many model simulations and large amounts of computer time, especially when the spatial domain is large or global. Furthermore, truly objective hand tuning in the full parameter space of a complex global land model is effectively impossible, thus the process usually relies on expert-informed tuning of a subset of parameters. However with this approach it is difficult to assess the uniqueness of the solution or how close the solution is to the optimal one. A machine learning approach to model calibration can provide increased computational efficiency and reduced analysis time as well as objective methods to assess calibration results. Here we use a neutral network-based global emulator of the Community Land Model, version 5, to calibrate parameter values informed by carbon and water flux observations using Bayesian inference. Parameter posterior distributions are produced using a Markov Chain Monte Carlo approach. By sampling from these distributions and running future climate simulations, we estimate the contribution of land model parameter uncertainty in future projections of climate change.