A new machine learning framework called PhyxMamba can reconstruct the behavior of chaotic systems, like turbulent fluids or brain activity, from just a handful of short, noisy observations.
Researchers built PhyxMamba by combining Mamba-based state-space models, a newer alternative to transformers, with physics-informed constraints. The system uses time-delay embeddings to rebuild the underlying "attractor" - the geometric shape describing how a chaotic system evolves over time - even when available data is sparse and disconnected. In tests on simulated and real-world chaotic systems, including the Lorenz96 climate model, PhyxMamba beat the strongest existing baseline by more than 44% in prediction accuracy and 8% in topological fidelity. It also held up against partial observations and noisy input, conditions that typically break older methods.
That data-scarcity problem is not a minor inconvenience. Chaotic systems show up in the fields scientists care most about - climate, neuroscience, fluid dynamics - and in all of them, long, clean recordings are expensive or outright impossible to collect. Reservoir computing, the prior go-to approach, needs long synchronization windows just to locate where a system sits on its attractor, which is exactly the kind of data these fields rarely have.
Whether PhyxMamba generalizes from clean benchmark systems to messier real-world sensor streams is the open question this paper does not settle.