Research map / Probabilistic and scientific modeling / Dynamical systems

StreamPhy: Streaming Inference of High-Dimensional Physical Dynamics via State Space Models

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Data assimilation from sparse observations · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes StreamPhy, an SSM-based framework (attention observation encoder + HiPPO SSM + FT-FiLM decoder) for streaming reconstruction of physical fields from irregular sparse measurements, reporting lower VRMSE and 20-100x faster inference than diffusion-based SDIFT.

Problem

Real-time reconstruction of high-dimensional spatiotemporal physical fields from irregular, sparse, asynchronous streaming measurements; existing diffusion and functional tensor methods are offline, need full temporal observations, or are expensive at inference.

Approach

Combines an attention-based encoder that maps variable-size observation sets to fixed latent vectors, a HiPPO-LegS state-space model that updates online over irregular time intervals, and FT-FiLM, a feature-wise modulated functional tensor decoder claimed to be strictly more expressive than the functional Tucker model.

Key results

• StreamPhy (MH) achieves VRMSE 0.0935 on Turbulent Flow (uniform, rho=3%) versus 0.3248 for SDIFT, the best baseline shown.
• Inference on Turbulent Flow takes 0.0419s for StreamPhy versus 5.14s for SDIFT on an RTX 4090, roughly 100x faster.
• Ablation: removing the SSM raises Turbulent Flow VRMSE to 0.8537, and replacing FT-FiLM with FTM raises it to 0.1435 (uniform, rho=3%).

Surprising findings

• Removing the SSM causes VRMSE to degrade to near 0.85-0.91, far worse than any baseline, suggesting the encoder-decoder alone cannot handle streaming inputs.

Limitations

• The state transition matrix A is predefined (HiPPO-LegS) and its design space is unexplored.
• No explicit incorporation of physical laws or priors into the model.
• Evaluated on only three datasets with a limited set of sampling patterns.

Experimental setup

Datasets: Turbulent Flow Ocean Sound Speed Active Matter

Benchmarks: Turbulent Flow Ocean Sound Speed Active Matter

Metrics: VRMSE inference time (s) parameter count

Baselines: SDIFT Senseiver LRTFR OFTD CATTE

Builds on: HiPPO S4 Mamba functional Tucker model FiLM Senseiver

Closest papers by meaning

• Generating Full-field Evolution of Physical Dynamics from Irregular Sparse Observations (NeurIPS 2025 · similarity 0.90)
• Sparsely-Supervised Data Assimilation via Physics-Informed Schrödinger Bridge (NeurIPS 2026 · similarity 0.87)
• SCENT: Robust Spatiotemporal Learning for Continuous Scientific Data via Scalable Conditioned Neural Fields (ICML 2025 · similarity 0.86)
• Predicting partially observable dynamical systems via diffusion models with a multiscale inference scheme (NeurIPS 2025 · similarity 0.86)
• Gated Inference Network: Inference and Learning State-Space Models (NeurIPS 2024 · similarity 0.86)
• Sparse Diffusion Autoencoder for Test-time Adapting Prediction of Complex Systems (NeurIPS 2025 · similarity 0.85)