Chimera states are patterns where identical oscillators in the same network split into coexisting coherent (synchronized) and incoherent (desynchronized) groups. This paper proposes a practical way to detect and classify different chimera types without hand-tuned thresholds. The authors test their approach on a network of Rayleigh oscillators, a standard model that is known to produce many kinds of chimera patterns.

The method has two main parts. First, it uses a windowed Fourier analysis — in practice the Fast Fourier Transform (FFT) applied to short time windows — to extract local signal features for each oscillator. From each window the method estimates a baseline level, an amplitude, a phase, and a frequency. The authors refine the raw FFT output by fitting a parabola to the spectrum near its maximum and then doing a short nonlinear fit to the time series to improve amplitude and phase estimates.

Second, the method summarizes how those features change across neighboring oscillators using a normalized total variation, a simple measure of local spatial change. These summary values are fed into unsupervised clustering algorithms, such as k-means and Gaussian mixture models. Because the clustering works on those normalized variations, it can separate regions of parameter space that correspond to coherent dynamics, phase chimeras, and amplitude-mediated chimeras without the need for ad hoc thresholds.

The authors apply this procedure while varying coupling strength and coupling range in the Rayleigh network. They report that the combined Fourier-plus-clustering approach finds parameter regions with chimera states and further distinguishes different chimera types. They also argue that conventional metrics can fail to identify some regimes clearly, and that their Fourier-based features give a more direct and robust picture of what each oscillator is doing in time and space.