Early detection of Hopf bifurcations in a prototypical fluid system via deep learning of unbinarized recurrence plots

We explore whether deep learning of unbinarized recurrence plots (RPs) can provide early warning signals for the onset of limit-cycle oscillations. We consider a prototypical fluid system exhibiting either a supercritical or a subcritical Hopf bifurcation. From hot-wire velocity measurements, we reconstruct delay-coordinate trajectories and compute unbinarized RPs whose entries are the pairwise distances between reconstructed state vectors. Rather than operating directly on the raw time series, we use these RPs to train a convolutional neural network (ResNet-18) to regress a capped, normalized proximity-to-onset label, yielding a continuous estimate of the proximity to the Hopf point. Compared with established precursors-including the variance, lag-1 autocorrelation, dominant spectral-peak drift, and generalized Hurst exponent-the proposed framework provides a smoother and more reliable warning signal across different operating conditions, including previously unseen ones, without requiring ad hoc instability thresholds. Saliency analysis indicates that the network relies primarily on the evolving boundaries between regions of high and low recurrence, linking predictive performance to changes in the reconstructed phase-space geometry. These results demonstrate the potential of this model-free generalizable framework for the early detection of Hopf bifurcations in fluid flows and other dynamical systems.