Latent-Space Dynamics
Latent-space dynamics models evolve a compressed representation forward in time, giving fast surrogates for high-dimensional simulations.
Dynamics in a small space
Once a high-dimensional state is compressed to a few latent coordinates by a learned or linear encoder, the remaining task is to predict how those coordinates change over time. A latent-space dynamics model does this, evolving the compact representation forward and decoding back to the full state only when output is needed. The heavy lifting happens in the small space.
Ways to model the evolution
- Neural ordinary differential equations that learn a continuous latent flow
- Recurrent networks that step the latent state discretely
- Sparse regression that finds a compact symbolic latent law
- Linear operators, as in Koopman and dynamic mode decomposition approaches
Stability over long rollouts
The central challenge is staying accurate over many steps. A model trained to predict only the next step accumulates error and can diverge. Training on multi-step rollouts, penalizing long-horizon error, and encouraging a well-conditioned latent flow all improve stability. Some approaches favor a linear latent evolution precisely because its long-term behavior is predictable.
Encouraging simple dynamics
The latent space is not unique; the encoder can be shaped so the dynamics within it are as simple as possible. Learning the encoder and the dynamics jointly lets the model choose coordinates in which the evolution is nearly linear or sparse, which makes the dynamics model smaller, more stable, and easier to interpret.
Uses
Latent-space dynamics power fast surrogates for control, forecasting, and design exploration where a full solver is too slow to run in the loop. Their reliability is bounded by the regime the training data cover, so an accompanying uncertainty estimate and periodic refresh from full-order simulations keep the surrogate trustworthy as conditions move.