Machine learning for weather and climate are worlds apart
Watson-Parris, D
Abstract
Modern weather and climate models share a\ncommon heritage, and often even components,\nhowever they are used in different ways to answer\nfundamentally different questions. As such, attempts\nto emulate them using machine learning should\nreflect this. While the use of machine learning\nto emulate weather forecast models is a relatively\nnew endeavour there is a rich history of climate\nmodel emulation. This is primarily because while\nweather modelling is an initial condition problem\nwhich intimately depends on the current state of the\natmosphere, climate modelling is predominantly a\nboundary condition problem. In order to emulate the\nresponse of the climate to different drivers therefore,\nrepresentation of the full dynamical evolution of the\natmosphere is neither necessary, or in many cases,\ndesirable. Climate scientists are typically interested in\ndifferent questions also. Indeed emulating the steadystate climate response has been possible for many\nyears and provides significant speed increases that\nallow solving inverse problems for e.g. parameter\nestimation. Nevertheless, the large datasets, nonlinear relationships and limited training data make\nClimate a domain which is rich in interesting machine\nlearning challenges.