Seminar

Cloud processes are hugely important to the evolution of weather, as well as the climate-scale radiative properties of the Earth. However, weather and climate models suffer owing to incomplete and approximated physical representation of these processes within so-called "parameterizations". In some cases, it is reasonable to hope that data, both from real-world observations as well as detailed "reference" models, can improve our understanding of these processes and optimize their formulation in model parameterizations. The formalism of Bayesian inference provides tools that can constrain the value of uncertain parameter values and quantify remaining uncertainty using available data but requires that the solution is not prohibitively conditional on other sources of uncertainty, for example initial condition uncertainty or model structural uncertainties. In many cases these cannot be ignored, and systematic Bayesian inference may be inappropriate. I will present research using Bayesian inference and machine learning to constrain parameters and quantify their uncertainty in the NASA GISS ModelE climate model, as well as preliminary results from applying a similar technique to constrain microphysical parameters within a Large Eddy Simulation of a drizzling stratocumulus cloud. I will also present observational and modeling studies of cases where systematic inference may be inappropriate — isolated thunderstorms observed by radar and lightning mapping array in Houston Texas — where instead fundamental insights can be gained by comparing features observed and simulated using state-of-the-art numerical weather models. I will discuss prospects for future work, and potential for moving beyond current limitations.