Artificial neural networks are applied to perform nonlinear weather regime diagnostics for the ECMWF seasonal forecasting system. The focus is on possible nonlinear low-dimensional structure of the extratropical atmospheric flow (particularly the NAO and PNA). Primary diagnostic techniques are cluster analysis (mixture model clustering) and two different nonlinear extensions of principal component analysis (NLPCA) using neural networks, that is, Kramer's feedforward network technique and nonlinear kernel-PCA. In order to assess the realism of the ECWMF model our results are compared with those from observational data.