| Abstract: |
| First-order phase transitions such as explosive death can cause abrupt and irreversible suppression of collective oscillations, making their critical prediction important for preventing dynamical collapse in complex systems. In this talk, I will present two recent machine-learning approaches for predicting such transitions from oscillatory data alone. For low-dimensional coupled oscillators, we adopt a parameter-aware next-generation reservoir computing framework that embeds control parameters into nonlinear feature construction and enables accurate identification of critical points, hysteresis, and bifurcation structures in unseen regimes. For oscillator networks, we further introduce a parameter-aware parallel reservoir computing framework, where node-wise reservoirs preserve scalability while retaining sensitivity to parameter variations. The method accurately predicts explosive amplitude death and oscillation death in networks with different topologies, including nearest-neighbor ring, small-world, and random regular networks, while also reconstructing the corresponding spatiotemporal dynamics. Together, these results provide an efficient and scalable data-driven paradigm for forecasting first-order phase transitions in nonlinear dynamical systems and complex networks. |
|