Study finds GNNs and network measures fail to generalize stability in oscillator networks

A new paper by Christian Nauck, Michael Lindner, Nora Molkenthin, Jürgen Kurths, Eckehard Schöll, Jörg Raisch, and Frank Hellmann (published on arXiv, 57 pages, 7 figures, 32 tables) takes a critical look at the relationship between network structure and functional stability in complex oscillator networks. The researchers tested whether common network measures—such as degree, clustering, or centrality—can predict dynamic stability in both linear and non-linear oscillator models, as well as in realistic power grid topologies.

The results are sobering: the correlation between network measures and stability is highly sensitive to the specific ensemble of networks considered. For instance, simply changing the mean degree from 6 to 8 can invert the sign of the correlation. This means that a measure that seems to predict instability in one set of networks might predict stability in another, making it unreliable for general use. The authors also tested machine learning approaches, including Graph Neural Networks (GNNs) and non-linear combinations of many network measures. Both techniques could accurately predict stability when trained and tested on networks from the same ensemble—but they failed when applied to a different ensemble, even if the difference was subtle.

The study's conclusion is stark: neither traditional network measures nor advanced ML models can reliably identify the structural causes of instability across varying network configurations. This challenges a core assumption in network science—that structural features can universally predict emergent dynamics. For applications like power grid stability, this means that models trained on one set of grid configurations (e.g., a specific topology or load pattern) might not transfer to others. The paper calls for more cautious use of both network measures and ML in stability analysis, and suggests that future work must account for ensemble dependence.