Understanding and predicting critical transitions in complex systems—such as neuronal networks—requires the ability to reconstruct their underlying dynamics and interaction structures directly from data. In this talk, I present a data-driven framework that combines theoretical model reduction with machine learning methods to tackle this challenge, particularly in weakly coupled chaotic networks. A key innovation of our approach lies in extracting information from the stochastic fluctuations typically dismissed as noise. These fluctuations, in fact, encode valuable details about the network’s structure and dynamics. We reconstruct effective network models that merge local dynamical rules with statistical representations of interactions. Applied to synthetic data resembling cat cortical networks, our method demonstrates the ability to forecast critical transitions even in unobserved parameter regimes [1]. Importantly, we also show that, under reasonable assumptions, it is possible to recover the full network dynamics using relatively short and sparse datasets. This reduces reliance on long-term observations or small system sizes, which are often impractical in real-world experiments. Extending this method to experimental neuronal data from the mouse neocortex, we demonstrate its potential for learning both the dynamic rules and the network topology [2]. This paves the way for reliable detection of critical regime shifts using limited information. I will conclude by discussing possible reconstruction challenges and our proposed solutions to this broader problem [3].