Projects‎ > ‎


Supermodeling by combining imperfect models (SUMO)

A supermodel is an interconnected ensemble of existing imperfect models of a real, observable system. The connections between the models can be learned from observational data using methods from machine learning. The supermodel outperforms the individual models in simulating the behaviour of the real system since it has learned to combine the strengths of the individual models. The concept of supermodeling is based on a new combination of insights from climate science, nonlinear dynamical systems, and machine learning.

For the simple example of Lorenz systems the supermodel accurately reproduces the true solution after learning and illustrates the potential of the supermodeling approach. The imperfect models are connected in such a way that the true evolution is simulated quite accurately. 

Figure 1. The solution for the true Lorenz equations is plotted in green in both panels. In the left panel the red trajectory denotes the solution of the supermodel with connections set to unity and in the right panel with connections learned from data.

By bringing together experts from the fields of climate science, non-linear dynamical systems and machine learning, we believe we are in an excellent position to develop a strategy that by the end of the project leads to a supermodel consisting of an ensemble of interconnected state-of-the-art climate models with improved simulation accuracy as compared to the current standard, multimodel ensemble mean approach. 

The project web site is: