CS-SHRED : Enhancing SHRED for Robust Recovery of Spatiotemporal Dynamics

Romulo Brito da Silva, Diego Passos, Cassio Machiaveli Oishi, JN Kutz · 2025

Read CS-SHRED : Enhancing SHRED for Robust Recovery of Spatiotemporal Dynamics by Romulo Brito da Silva, Diego Passos, Cassio Machiaveli Oishi, JN Kutz (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-22303.txt.

Introduction

Da Silva and co-authors present a 2025 paper on rebuilding full motion fields from sparse or damaged sensor data. Their base method, SHRED, uses a shallow recurrent decoder to fill in space-time patterns. Real data, though, can be noisy, gappy, or pressed into few numbers. The new design, CS-SHRED, folds compressed sensing tricks into the decoder.

A penalty term favors simple signals, which helps recovery from few sensors. A smart loss mixes two error types and adapts to local noise levels, guarding fine detail. Tests span sticky fluid flows, humidity fields, sea surface heat, and swirling turbulence. Scores beat the base method on sharpness and error, the authors report.

Catalog record

CREATORRomulo Brito da Silva, Diego Passos, Cassio Machiaveli Oishi, JN Kutz
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE89 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-22303

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

CS-SHRED : Enhancing SHRED for Robust Recovery of Spatiotemporal Dynamics Thanks: Citation : da Silva R. B., Passos, D.; Oishi, C. M.; Kutz, J.N. CS-SHRED: Enhancing SHRED for Robust Recovery of Spatiotemporal Dynamics. Romulo Brito da Silva, Diego Passos, Cassio Machiaveli Oishi, JN Kutz Abstract We present CS-SHRED , a novel deep learning architecture that integrates Compressed Sensing (CS) into a Shallow Recurrent Decoder ( SHRED ) to reconstruct spatiotemporal dynamics from incomplete, compressed, or corrupted data. Our approach introduces two key innovations. First, by incorporating CS techniques into the SHRED architecture, our method leverages a batch-based forward framework with ℓ 1 \ell_{1} regularization to robustly recover signals even in scenarios with sparse sensor placements, noisy measurements, and incomplete sensor acquisitions. Second, an adaptive loss function dynamically combines Mean Squared Error (MSE) and Mean Absolute Error (MAE) terms with a piecewise Signal-to-Noise Ratio (SNR) regularization, which suppresses noise and outliers in low-SNR regions while preserving fine-scale features in high-SNR regions. We validate CS-SHRED on challenging problems including viscoelastic fluid flows, maximum specific humidity fields, sea surface temperature distributions, and rotating turbulent flows. Compared to the traditional SHRED approach, CS-SHRED achieves significantly higher reconstruction fidelity—as demonstrated by improved SSIM and PSNR values, lower normalized errors, and enhanced LPIPS scores—thereby providing superior preservation of small-scale structures and increased robustness against noise and outliers. Our results underscore the advantages of the jointly trained CS and SHRED design architecture which includes an LSTM sequence model for characterizing the temporal evolution with a shallow decoder network (SDN) for modeling the high-dimensional state space. The SNR-guided adaptive loss function for the spatiotemporal data recovery establishes CS-SHRED as a promising tool for a wide range of applications in environmental, climatic, and scientific data analyses. K eywords Spatiotemporal dynamics, Compressed sensing, CS–SHRED, Sparse/incomplete measurements, LSTM sequence modeling, Spatiotemporal field reconstruction 1 Introduction The SHRED (SHallow REcurrent Decoder) [ 1 ] framework offers a novel approach to reconstructing spatiotemporal dynamics using multivariate time series from sensor measurements. Recognized for its data-driven nature, SHRED demonstrates robustness even when sensors are mobile [ 2 ] or not optimally placed. Indeed, the SHRED architecture leverages the Taken’s embedding theorem and separation of variables in order to produce stable spatio-temporal reconstructions and forecasts from minimal sensor measurements. The SHRED architecture processes multivariate sensor time series data through Long Short-Term Memory (LSTM) recurrent layers. A fully connected, shallow decoder receives the latent representation from the LSTM, aiding in reconstruction and restoration to the high-dimensional state. Across various datasets—including isotropic turbulent flow, sea surface temperature, and atmospheric ozone concentration— SHRED outperforms existing linear and nonlinear methods [ 1 ] . Remarkably, SHRED achieves accuracy comparable to methods requiring a larger number of optimally placed sensors, even with just three randomly placed sensors for 2D spatio-temporal data. Furthermore, SHRED provides real-time compression, independent of the underlying physics, for modeling engineering and physical systems with low input requirements. SHRED can be used for reduced order modeling [ 3 ] , model discovery [ 4 ] and digital twin architectures [ 5 ] . Although robust to noisy sensor measurements, the SHRED model shows limitations in handling spatiotemporal data with gaps in the sensor measurements, corrupt measurements, or insufficient sampling at various time points. These shortcomings become evident when employing a methodology designed to simulate adverse conditions of information loss during sensor data acquisition. In our approach, Compressed Sensing (CS) is employed to mitigate the limitations of practical data-acquisition systems. Specifically, we perform random subsampling that (i) removes a fixed percentage of spatial samples and (ii) applies this removal only to a randomly selected subset of snapshots. As a result, the remaining measurements come from only a few sensors, sparsely positioned and suboptimally distributed across the domain, producing sparse and irregularly subsampled spatiotemporal series. This process mimics scenarios in which data are collected incompletely or inadequately due to various real-world factors. In practical applications, data collection often experiences information loss due to sensor malfunctions, limited coverage, or environmental disturbances. For example, environmental monitoring measurements can be compromised by adverse weather, interference from nearby objects, or sensor technical issues. Similarly, in medical imaging, artifacts or patient movement during scanning may result in incomplete or distorted data. Thus to address these limitations, we propose integrating the Compressed Sensing (CS) framework into SHRED , creating CS-SHRED (Compressed Sensing SHRED). This integration enhances the model’s ability to recover missing or incomplete data points while maintaining SHRED ’s efficiency in reconstructing spatiotemporal dynamics from a limited number of sensors. The approach introduces a restriction operator that confines the signal to observed (non-null) elements, focusing the recovery problem on known data. The formulation combines both the restriction operator and a sparsifying operator within a sparse minimization framework using convex relaxation, facilitating the accurate recovery of the signal captured by the sparse randomly placed sensors. Integrating SHRED with the unsupervised learning framework of CS significantly advances the recovery of spatiotemporal dynamics using a limited number of sensors without requiring prior knowledge of the governing laws. Additionally, we develop a novel loss function that incorporates regularization terms related to the signal-to-noise ratio (SNR). This loss function combines traditional L1 and L2 regularization with an SNR-based term, allowing the model to prioritize areas with higher SNR and penalize those with lower SNR. By focusing on more reliable data, the model’s overall reconstruction performance is improved, enhancing its resilience to noise and data imperfections. The ability to handle missing data is particularly relevant in fields such as exploration seismology, where data gaps can compromise interpretation and the identification of oil and gas reservoirs [ 6 , 7 ] . Similarly, in medical imaging, such as magnetic resonance imaging (MRI), missing information may distort images and impact medical diagnosis [ 8 ] . To address these challenges, CS has been widely employed as an effective technique for recovering missing information in subsampled or corrupted data. CS leverages the concept that many real-world signals are naturally sparse or can be approximated by a small number of coefficients in a suitable basis. By reconstructing missing data using CS, it is possible to achieve satisfactory results in various applications, considering certain limitations and conditions. This contributes to the accurate analysis and interpretation of data in many contexts. Despite its advantages, CS also has certain limitations that need to be considered. One of the main challenges is the high computational cost associated with the optimization process in the l_{1} norm, particularly in large datasets. This procedure may require substantial computational resources and prolonged execution time, making it impractical in certain scenarios [ 9 ] . Furthermore, the performance of CS can be influenced by the quality and representativeness of the input data. If the data do not adhere to the assumption of sparsity or are not properly preprocessed, CS may yield suboptimal or inadequate results [ 10 ] . Another drawback is the sensitivity of CS to certain configuration parameters, such as the size of the dictionary used in the sparse representation of the data. Improper selection of these parameters can lead to poor data reconstruction, compromising the usefulness and reliability of the results [ 11 ] . In addition to Compressed Sensing , various other techniques have been explored for recovering missing

Read the complete text