Improving simulations by combining imperfect models through learning

Abstract

With the uprising climate problems like the warming due to greenhouse gasses, the need for future climate scenarios is growing fast. Many state of the art climate models exist, but they are imperfect and give a large variaty of future scenarios. In this thesis we try to find a way to improve the predictions of climate models by combining them. Multiple imperfect models will exchange information during the simulations and are combined into a super-model. The idea is not only that the imperfect models can use each others strengths, but also that synchronization between the models can occur, such that a mutual prediction results. The objective is to find a way to let models exchange information such that the resulting super-model gives a better prediction than any of the separate imperfect models. A learning process is developed to objectively determine the exchange of information between the models. The approach is tested on small chaotic dynamical systems that have similar properties as the atmosphere. The system with standard parameter values is regarded as truth and three imperfect models are created by perturbing these parameter values. By using the small chaotic dynamical systems we can use the information of the truth to see how well the approach works.