Abstract
In this paper, a variety of machine learning models for reducing climate model inaccuracy are developed and critically examined. The most effective model at mitigating climate model inaccuracy is a random forest regressor, which reduces temperature Root Mean Square Error (RMSE) from 2.90 to 0.44 in the Global Ensemble Forecast System (GEFS). Multiple linear models, neural networks, and random forest regressor correction models are trained on a large climate model inaccuracy dataset. This inaccuracy dataset is created by comparing the results of a climate reanalysis with the results of a climate reforecast, assuming that the reanalysis is more accurate at representing real climate values than the reforecast. This assumption is successfully validated by comparing both datasets to an observational validation set. The random forest correction model performs significantly better than the other correction approaches, for which possible explanations are discussed. Finally, this method of climate model correction is applied to a generalized setting, creating a program that can automatically generate a tailor-made random forest correction model for any climate model output.