Physics-based learning models for future climate predictions

Weymouth, G.D. and Grandey, B. (2012) Physics-based learning models for future climate predictions. Frontiers in Computational Physics: Modeling the Earth System. 16 - 20 Dec 2012.

Abstract

Modern climate projection largely relies on state-of-the-art atmosphere-ocean general circulation models (AOGCMs). However, the size and complexity of these models makes them extremely computationally expensive. In order to explore the parameter space of future climate uncertainty, the AOGCMs are supplemented in practice by simplified climate models, which have significantly shorter evaluation times. However, this comes at the cost of capturing only a small subset of the relevant physics, and the data produced by the simplified models often contain biases compared to the AOGCMs. A common practice is to `tune’ the simple models by adding non-physical internal coefficients that are then fixed using AOGCM data. However, this is a nonlinear programming problem; it is error prone, time intensive, and often inaccurate. Here, a novel machine learning technique, physics-based learning models (PBLM), is applied to rigorously combine AOGCM data and simplified climate models in order to produce an improved estimate of the climate uncertainty space. PBLM provides a general framework for incorporating simplified physics-based models (which capture fundamental aspects of the climate system) into modern machine learning regression methods, greatly improving their accuracy and reducing their dependence on costly example data from the AOGCM. The application of PBLM to simplified global climate models using a small set of example AOGCM projections demonstrate the speed, simplicity, and accuracy advantages of the new approach

This record has no associated files available for download.

Contributors

Author: G.D. Weymouth

Download statistics