Skip to main content
Back to timeline
arXivSource publication:

A few point sensors can reproduce chaos: a latent time stepper recovers the Kuramoto–Sivashinsky Lyapunov spectrum

Synopsis

Sparse point measurements alone can yield latent dynamics that advance autonomously. A shallow recurrent decoder (SHRED) encodes a history of measurements into a latent state, and an echo state network learns to evolve that state on its own, bypassing the step of reconstructing the full field before encoding it. On the Kuramoto–Sivashinsky equation the autonomous latent model reproduces the leading Lyapunov exponents of the true system; on Kolmogorov flow it stays on the latent manifold traced by the encoded trajectory over long horizons.

Source-provided article image: Replicating chaotic dynamics from sparse sensors with recurrent latent states
arXiv

Interpretation

The latent state can be built directly from sparse sensor measurements and advanced autonomously, without first reconstructing and then compressing the full field. Reduced-order models usually take the full state as encoder input, so a sensor-limited deployment must first solve an ill-posed inverse problem, and that error enters the latent representation. On the Kuramoto–Sivashinsky equation the encoder receives only a fraction of the state components as input, yet the autonomous latent model reproduces the leading Lyapunov exponents of the true system.

A shallow recurrent decoder (SHRED) encodes a history of point measurements into the latent state, and an echo state network then learns an autonomous map on that latent sequence. The SHRED encoder is a gated recurrent unit that reads a time-lagged measurement window rather than a single snapshot, so no full state is needed at deployment. Only the linear readout of the echo state network is trained, so the Jacobian of its closed-loop map is available in closed form and the Lyapunov spectrum of the learned dynamics can be computed directly.

Reconstruction accuracy does not determine forecastability: the recurrent encoding that reconstructs worse forecasts better on the more complex chaotic case. At equal latent dimension SHRED reconstructs less accurately than a convolutional autoencoder given the full state, though far more accurately than a POD projection at the same dimension. The two pipelines forecast comparably on the simpler KS case, and SHRED plus the echo state network forecasts better on the more complex KS case even though the autoencoder's reconstruction error there is about six times smaller.

Perspective

The demonstrated scope is the Kuramoto–Sivashinsky equation at two domain lengths and Kolmogorov flow at a single Reynolds number where the flow is quasi-periodic; the latter tests mainly sensing and reconstruction, and its long-time fidelity rests on a qualitative comparison of latent manifolds.

Forecast accuracy drops from the smaller to the larger domain, which the authors raise as a question of how the approach scales to higher-dimensional attractors, without establishing the sensor budget needed to sustain spectrum replication; Kolmogorov flow is studied at one quasi-periodic Reynolds number, and higher-Reynolds turbulent cases remain untested.