LAtent Phase Inference from Short time sequences using SHallow REcurrent Decoders (LAPIS-SHRED)

Abstract: Reconstructing full spatio-temporal dynamics from sparse observations in both space and time remains a central challenge in complex systems, as measurements can be spatially incomplete and can be also limited to narrow temporal windows. Yet approximating the complete spatio-temporal trajectory is essential for mechanistic insight and understanding, model calibration, and operational decision-making. We introduce LAPIS-SHRED (LAtent Phase Inference from Short time sequence using SHallow REcurrent Decoders), a modular architecture that reconstructs and/or forecasts complete spatiotemporal dynamics from sparse sensor observations confined to short temporal windows. LAPIS-SHRED operates through a three-stage pipeline: (i) a SHRED model is pre-trained entirely on simulation data to map sensor time-histories into a structured latent space, (ii) a temporal sequence model, trained on simulation-derived latent trajectories, learns to propagate latent states forward or backward in time to span unobserved temporal regions from short observational time windows, and (iii) at deployment, only a short observation window of hyper-sparse sensor measurements from the true system is provided, from which the frozen SHRED model and the temporal model jointly reconstruct or forecast the complete spatiotemporal trajectory. The framework supports bidirectional inference, inherits data assimilation and multiscale reconstruction capabilities from its modular structure, and accommodates extreme observational constraints including single-frame terminal inputs. We evaluate LAPIS-SHRED on six experiments spanning complex spatio-temporal physics: turbulent flows, multiscale propulsion physics, volatile combustion transients, and satellite-derived environmental fields, highlighting a lightweight, modular architecture suited for operational settings where observation is constrained by physical or logistical limitations.

• The paper introduces LAPIS-SHRED, a modular architecture that combines SHRED spatial encoding with simulation-trained latent temporal models and a frozen decoder to infer complete trajectories from as few as three sensors and 7% of the observation period.
• Across six physical systems, the method achieves NRMSE values as low as 0.025 and below 5% on several chaotic benchmarks, while supporting both forward forecasting and backward reconstruction from terminal observations.
• The results show strong data efficiency but also expose limits: uncertainty quantification, genuine spatial domain shift, chaotic-system theory, active sensing, and autoregressive error accumulation remain open challenges.

LAPIS-SHRED addresses a specific and severe observational regime in scientific machine learning: reconstructing or forecasting complete spatiotemporal trajectories when sensor measurements are both hyper-sparse spatially (as few as 3 point sensors) and confined to short temporal windows (as little as 7% of the trajectory, or a single terminal frame). The architecture builds on the SHallow REcurrent Decoder (SHRED) family (Maity et al., 2024) and adds a latent-space temporal inference stage trained exclusively on simulation data. This essay summarizes the method, its empirical evaluation across six physical systems, the supporting theoretical analysis, and its limitations.

Motivation and problem setting

The paper targets settings where observations are unavailable over most of the temporal domain: post-event deposits in geophysics, post-mortem structural damage, early-phase combustion monitoring, and brief satellite revisit windows. Formally, given sparse sensor observations $\mathbf{s}_t = \mathbf{M}\mathbf{Y}_t$ with $p \ll n$ over only $W+1$ frames ($W \ll T$), the task is to recover all $T$ full spatial states. Two complementary regimes are considered: backward inference from terminal windows (reconstructing antecedent dynamics) and forward inference from initial windows (forecasting). The authors argue that classical data assimilation requires dense sequential observations, PINNs require known PDE forms and anchoring constraints, neural operators such as FNO (Li et al., 2020) and DeepONet (Wang et al., 2021) require large paired full-domain datasets, and POD/DMD-based ROMs cannot handle hyper-sparse sensing—leaving SHRED-based architectures as the natural foundation.

Architecture

LAPIS-SHRED is a three-stage modular pipeline:

1. Pre-trained SHRED spatial module: an LSTM temporal unit maps lagged sparse sensor histories to a low-dimensional latent code $\mathbf{z}_t \in \mathbb{R}^{d_z}$, decoded by a shallow MLP to the full spatial state. The module may operate frame-by-frame (preferred for chaotic systems, leveraging Takens' embedding), in Seq2Seq mode (for dissipative systems with global temporal coherence), or as a multi-scale LF/HF decomposition (Cheap2Rich/SENDAI-style).
2. Temporal dynamics model: trained on simulation-derived latent trajectories to propagate latents forward or backward. Two variants are provided—a Seq2Seq model that emits the entire unobserved latent trajectory in one pass (avoiding autoregressive error accumulation, at the cost of fixed output length), and an autoregressive BiLSTM+MLP model for open-ended horizons (with coupled LF-conditioned HF rollouts in multi-scale settings).
3. Frozen decoder: predicted latents are mapped to spatial fields by the frozen decoder, providing regularization toward the decoder's representational space and enabling modular replacement of components.

Two design details merit emphasis. For the extreme single-frame case, a static padding strategy replicates the terminal sensor vector $L$ times; training data are augmented with analogous padded sequences so the temporal unit learns to interpret constant signals as stationarity. Training uses ensembles of $K$ simulation trajectories spanning parameter configurations (or sub-sequences of a single long trajectory for temporal diversity), so no ground-truth observations beyond the short window are needed at deployment.

Empirical results

Six experiments span chaotic PDEs, vortex-dominated flows, multiscale propulsion physics, combustion transients, and satellite-derived environmental fields. Key quantitative outcomes:

Experiment	Direction	Observed window	Sensors	NRMSE
2D Kuramoto–Sivashinsky	Backward	~10%	3	0.046
2D Kolmogorov flow	Backward	~10%	8	0.036–0.044
2D von Karman vortex street	Fwd / Bwd	10%	5	0.037 / 0.033
High-fidelity RDE	Forward	25 frames	25	0.114
1D RDE ignition	Backward	20%	16	0.025 (P) / 0.032 (T)
MODIS NDSI snow cover	Fwd / Bwd	7% / single frame	64	0.167 / 0.130

Several results stand out. On the chaotic benchmarks, LAPIS-SHRED achieves NRMSE below 5% despite positive Lyapunov exponents precluding long-horizon deterministic prediction—the simulation prior plus short observation window provide sufficient regularization. On the high-fidelity rotating detonation engine dataset (a simulation costing over 2 million CPU-hours per operating condition), autoregressive rollout maintains reconstruction quality matching the Cheap2Rich baseline (RMSE 0.108 vs. 0.101) and produces 200 physically plausible extrapolation steps beyond the data horizon, with the LF detonation wavefronts remaining coherent while HF fluctuations attenuate—an interpretable signature of error accumulation. On NDSI, backward inference from a single terminal frame outperforms forward inference from five frames (NRMSE 0.130 vs. 0.167), which the authors attribute to the stronger constraint imposed by the near-complete-melt terminal state versus the stochasticity of melt events.

An ablation study on NDSI isolates each component: errors saturate beyond roughly 32 sensors; performance stabilizes for SHRED hidden dimension $\geq 32$; the temporal model exhibits a capacity–generalization trade-off peaking at $d_h^B = 64$ for forward inference; observation-window length improves performance monotonically up to $p \ll n$0, consistent with Takens' theorem; and static padding length has an interior optimum at $p \ll n$1.

Baseline comparisons

Against SHRED-ROM [2506.xxxx, Tomasetto et al.]—the most directly comparable method, which lacks a dedicated temporal dynamics model—LAPIS-SHRED performs comparably on HF-RDE but substantially better on NDSI forward prediction (NRMSE 0.167 vs. 0.191–0.214), while uniquely supporting backward inference, which SHRED-ROM's forward-only LSTM processing cannot accommodate. Against broader operator-learning methods, the paper argues structurally rather than empirically: standard FNO requires full spatial fields, RecFNO-type extensions lack temporal propagation, S-DeepONet assumes temporally dense loadings, and neither framework natively supports backward-time inference. A capability table shows LAPIS-SHRED as the only evaluated method supporting sparse spatial sensing, temporal sequence encoding, and bidirectional inference simultaneously. Notably, the paper concedes that even on the easier reconstruction-only task, SHRED-ROM already outperforms PDS, POD-AE-SE, and POD-DeepONet, so the comparison set does not include methods specifically designed for this regime.

Error analysis

A four-term decomposition separates total error into encoding, temporal-model, decoding, and sim-to-real contributions. The formal guarantees are developed only for backward inference in dissipative systems converging to stable equilibria (the NDSI-like case): naive inversion of a dissipative flow amplifies terminal noise as $p \ll n$2, but restricting the inverse to the low-dimensional manifold spanned by the simulation ensemble bounds the effective sensitivity. The resulting bound scales with the Lipschitz constant of the learned backward map, the encoding error (which grows linearly with padding length $p \ll n$3), and the off-manifold sim-to-real gap. This yields an explicit bias–variance trade-off in ensemble rank $p \ll n$4: larger $p \ll n$5 reduces off-manifold error but admits faster-decaying modes with larger $p \ll n$6. The paper is explicit that this analysis does not extend to chaotic regimes or to autoregressive forward rollout with cumulative error.

Limitations and open questions

The authors identify several limitations directly. The framework produces point estimates without calibrated uncertainty quantification, which they deem essential for operational decision support. The observation window length is fixed a priori rather than chosen by active sensing. Theoretical error bounds cover only dissipative equilibria; chaotic-regime information-theoretic constraints and autoregressive error accumulation remain uncharacterized. The NDSI experiment eliminates spatial domain shift by construction (identical terrain across years), so the sim-to-real gap tested there is limited to inter-annual meteorological variability—a narrower challenge than genuine cross-geometry deployment. Finally, per-system training remains required; whether a shared latent-space prior across PDE families (multi-physics pretraining) can replace it is left open.

Conclusion

LAPIS-SHRED extends the SHRED family with a latent-space temporal inference stage that recovers full spatiotemporal trajectories from observation windows as short as 7% of the temporal domain—or a single terminal frame—using as few as 3 sensors, with NRMSE consistently within 15% of a full-observation SHRED baseline across six diverse physical systems. Its principal strengths are architectural modularity (interchangeable spatial modules, frozen decoders, bidirectional inference) and extreme data efficiency at deployment, underwritten by simulation-only training. The main open problems are uncertainty quantification, extension of the error theory beyond dissipative equilibria, and validation under genuine spatial domain shift.