Smith, L.A. Proc. National Academy of Science (2001)
Abstract
Climate models are, for the most part, large dynamical systems involving a million (or more) variables on big computers. Given that they are nonlinear and that they are not perfect, what can we expect to learn from them about the Earth's climate? How can we determine which aspects of their output might be useful and which are noise? and how should we distribute resources between making them ''better'', estimating variables of true social and economic interest, and quantifying how good they are at the moment? Just as ''chaos'' prevents accurate weather forecasts, so model error precludes accurate forecasts of the distributions which define climate. This yields uncertainty of the second kind. Can we estimate the uncertainty in our uncertainty estimates? These questions are discussed. Ultimately, all uncertainty is quantified within a given modelling paradigm; our forecasts need never reflect the uncertainty in a physical system.