Quantifying Output Uncertainty in Models used for Climatic Change Research Seminar - Artificial neural network assisted Bayesian calibration of Earth Systems Models
This is part of the Quantifying Output Uncertainty in Models used for Climatic Change Research Seminars.
How can we tractably and rigorously combine data from observations and computationally expensive earth system models/simulators to infer past and future climate/earth system evolution with appropriate uncertainty estimation? I will present an evolving methodology that relies on brute force Markov Chain Monte Carlo sampling to generate a posterior distribution for model parameters given observational constraints. Bayesian artificial neural network emulators of the simulator provide computational tractability for such sampling. Through two concrete examples (reconstruction of deglacial ice sheet evolution and general circulation climate model calibration), I will illustrate the strengths and ongoing challenges in the application of the methodology, especially within the context of trying to quantify the structural contribution to uncertainty.
Contact email@example.com for more information about this event.