Machine learning topological invariants of non-Hermitian systems

The study of topological properties by machine learning approaches has attracted considerable interest recently. Here we propose the invariants that are unique in non-Hermitian systems. Specifically, train neural networks to predict winding eigenvalues four prototypical Hamiltonians on complex energy plane with nearly $100\%$ accuracy. Our demonstrations Hatano-Nelson model, Su-Schrieffer-Heeger model and generalized Aubry-Andr\'e-Harper one dimension, two-dimensional Dirac fermion terms show capability exploring associated phase transitions diagrams Moreover, trained a small data set diagram can successfully untouched regions. Thus, our work paves way revealing topology toolbox.