Turbulence theory by machine

The question of whether especially significant sub-volumes of a turbulent flow can be `blindly' identified by automatic means, independently of a-priori assumptions, is addressed using the example of two-dimensional decaying turbulence. Significance is defined as influence on the future evolution of the flow, and the problem is cast as an unsupervised machine `game' in which the rules are the Navier--Stokes equations. It is shown that significance is an intermittent quantity in this flow, and that it is different from relevance. For example, high-energy regions are not necessarily more significant than low energy ones. In accordance with previous intuition, the most significant features are found to be vortices, while the least significant ones are dominated by strain. Subject to cost considerations, the method should be applicable to more general turbulent flows.