Using Machine Learning to Accelerate MCMC for the CARbon DAta-MOdel fraMework
Abstract
Understanding the terrestrial carbon and water cycles and their uncertainties is key for future climate projections. Terrestrial biosphere models are increasingly sophisticated, however these generally remain largely unchallenged by a growing volume of in-situ, tower, aircraft and satellite observations. Bayesian Model-data integration techniques are key for inferring unknown or uncertain mechanistic model parameters and states from the growing Earth observation record. Furthermore, Markov Chain Monte Carlo (MCMC) approaches can be used in combination with model-data fusion frameworks to extract parameter states and uncertainties; however, computational costs involved can be prohibitive for many applications. To reduce the computational burden we evaluate the possibility of using machine learning to accelerate MCMC for the CARbon DAta-MOdel fraMework (CARDAMOM). Instead of emulating the highly non-linear model itself we reduce the complexity by directly predicting the likelihood that a set of ecosystem parameters agrees with observations. We present challenges and possibilities of such an approach.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.B54A..05L