Séminaire : Mariana Clare (Imperial College): Assessing uncertainties in atmosphere and ocean models
Tuesday 30 November 2021, 02:00pm - 03:00pm
Hits : 13
Abstract: Ocean and atmosphere models suffer as decision support tools due to the high degree of uncertainty associated with them, both due to incomplete knowledge of data and natural variability in the system. These uncertainties can be assessed using a range of statistical techniques. A traditional technique to assess these uncertainties is the Monte Carlo method. However, the latter can be computationally unfeasible to implement because it often requires thousands, if not millions, of computationally expensive model simulations. In the first part of this talk, I will outline how multilevel Monte Carlo methods and multilevel multifidelity Mote Carlo methods can significantly reduce computational cost whilst maintaining accuracy. I will focus, in particular, on uncertain parameters in coastal ocean models. In recent years, machine learning techniques have also begun to be used to assess uncertainty. In the second part of my talk, I will show how neural networks can be used to predict full probability density functions for output variables. I will focus, in particular, on the field of weather forecasting, and show that the key probabilistic information provided by my neural network approach is vital for making informed weather forecasts.