Neural-network maps for two-parameter modeling of bistability and codimension-two bifurcations in two-dimensional flow dynamical systems.

We develop a machine-learning approach to reproduce the behavior of two versions of the van der Pol oscillator exhibiting a subcritical Andronov-Hopf bifurcation, with or without a codimension-2 Bautin point. We construct a neural-network model that functions as a recurrent map and train it on short segments of oscillator trajectories. The results show that, despite its relatively simple architecture of neural network, the trained map is capable of capturing the key dynamical features of the original oscillators, even though corresponding examples are absent from the training data. In a two-dimensional parameter space, the neural-network map reproduces, without direct exposure during training, the multistability of the oscillators and their bifurcations, including the Andronov-Hopf bifurcation, the saddle-node bifurcation associated with the birth of a limit cycle, and the codimension-2 Bautin bifurcation. We further show that, although explicit examples of these effects are not necessary, the ability to recover them still depends on an appropriate choice of training data.