Characterizing Exceptional Points Using Neural Networks

One of the key features of non-Hermitian systems is the occurrence of exceptional points (EPs), spectral degeneracies where the eigenvalues and eigenvectors merge. In this work, we propose applying neural networks to characterize EPs by introducing a new feature -- summed phase rigidity (SPR). We consider different models with varying degrees of complexity to illustrate our approach, and show how to predict EPs for two-site and four-site gain and loss models. Further, we demonstrate an accurate EP prediction in the paradigmatic Hatano-Nelson model for a variable number of sites. Remarkably, we show how SPR enables a prediction of EPs of orders completely unseen by the training data. Our method can be useful to characterize EPs in an automated manner using machine learning approaches.