Latent Phase Inference: Reconstructing Dynamical Systems from Temporal Fossils

The challenge of reconstructing complete spatiotemporal dynamics from fragmentary evidence sits at the intersection of inverse problems and scientific inference. In paleontology, researchers reconstruct entire ecosystems from isolated fossil fragments. In fluid dynamics and climate science, we face an analogous problem: recovering the full evolution of complex physical fields from spatially sparse sensors operating within severely constrained temporal windows. The paper LAtent Phase Inference from Short time sequences using SHallow REcurrent Decoders (LAPIS-SHRED) by Bao, Zhang, and Kutz presents a methodological framework that addresses this challenge through a modular, latent space approach. By separating the encoding of spatial information from the propagation of temporal dynamics, LAPIS-SHRED achieves what traditional methods cannot: accurate reconstruction of complete physical trajectories using as few as three sensors and observation windows comprising merely 7% of the total temporal domain.

The Architecture of Extreme Inference

LAPIS-SHRED operates through a three-stage pipeline that carefully partitions the inverse problem into manageable subcomponents. Stage one employs the SHRED (SHallow REcurrent Decoder) architecture, pretrained exclusively on simulation data to map hyper-sparse sensor time histories into a structured latent space. Unlike end-to-end approaches that attempt to directly map sparse observations to full field reconstructions across all time steps simultaneously, SHRED compresses the spatial complexity into a low-dimensional latent representation that preserves the essential dynamical features of the underlying physics.

Stage two introduces a temporal sequence model, specifically a Bidirectional Long Short-Term Memory (BiLSTM) network, trained on simulation-derived latent trajectories. This component learns to propagate latent states forward or backward in time, effectively bridging the unobserved temporal regions that separate short observation windows from the full trajectory. The bidirectional nature proves crucial; it allows the framework to perform both forecasting from initial conditions and backward reconstruction from terminal snapshots, a capability rarely found in standard operator learning frameworks.

At deployment, stage three combines these frozen, pretrained components. The system receives only a short observation window from the true physical system, compresses it through the fixed SHRED encoder, uses the temporal model to infer the complete latent trajectory, and finally decodes back to physical space. The architecture handles extreme cases, including single-frame terminal inputs where the observation window collapses to a single temporal slice, effectively reconstructing history from a solitary snapshot.

The technical results across six diverse experimental settings demonstrate remarkable robustness. From turbulent flows and multiscale propulsion physics to volatile combustion transients and satellite-derived environmental fields, LAPIS-SHRED consistently achieves low Normalized Root Mean Square Error (NRMSE) despite severe