The systems hardest to predict turn out to be the easiest for AI to learn from scratch.
Researchers publishing on arXiv have worked out a formal answer to a question that has nagged data-driven science for years: when can an AI actually recover a system's governing equations from observations alone? The answer hinges on chaos. Systems that exhibit chaotic behavior across their entire domain can, in principle, be uniquely identified from a single observed trajectory. The classical Lorenz system — a textbook example of chaos — is proven here to be analytically discoverable for the first time. Systems that are merely chaotic on a strange attractor can also be recovered, provided a specific geometric condition holds.
The flip side is the harder news. Stable, predictable systems — the kind engineers rely on for digital twins, robotics, and structural modeling — are often not discoverable from trajectory data alone. If a system has conserved quantities called first integrals, unique identification from data is mathematically impossible without additional prior knowledge baked in. That quietly undermines a lot of the confidence placed in purely data-driven modeling for engineering applications.
The finding reframes why weather forecasting has been such a productive target for machine learning: the atmosphere is chaotic, which turns out to be a feature for discoverability, not just a bug for prediction. For the tidy, well-behaved systems engineers actually want to simulate, more data is not the answer — the approach itself needs to change.