Can we estimate atmospheric predictability by performance of neural network forecasting? The toy case studies of unforced and forced Lorenz models

We present an analysis of the predictability for several regions on the attractor of the Lorenz-63 system, a simple nonlinear model which mimics some features of the atmosphere, like its chaotic behaviour and the presence of preferred states or "regimes". In this framework, through a forecasting activity on the attractor, a multilayer perceptron shows its ability to recognise different values of predictability in various zones of the attractor, if compared with other estimations of local predictability, like the growth rates of the so called "bred vectors". Furthermore, following recent studies on the impact of weak imposed forcings on the Lorenz model, as a toy simulation of increased anthropogenic forcings on the climate system, we analyse the changes of predictability for a new scenario by neural network forecasting. Therefore, even if the present paper must be considered as a preliminary attempt at the use of neural networks for predictability assessments, this activity shows good results and opens perspectives of further improvements and applications.