FR-Mamba: Time-Series Physical Field Reconstruction Based on State Space Model

Jiahuan Long, Wenzhe Zhang, Ning Wang, Tingsong Jiang, Wen Yao

TL;DR

This paper introduces FR-Mamba, a novel framework for reconstructing complex physical fields over time from sparse sensor data. By combining the long-range temporal modeling capabilities of State Space Models (SSMs) with the spatial feature extraction power of Fourier Neural Operators (FNOs), FR-Mamba achieves state-of-the-art accuracy in reconstructing unsteady flow fields, outperforming existing methods, especially for long sequences.

What Problem Does This Paper Address?

Accurately reconstructing the full spatial distribution of physical quantities (like velocity, pressure, or temperature) from limited sensor measurements is crucial for understanding and predicting complex systems, particularly in fluid dynamics and thermodynamics. Existing deep learning methods often struggle to capture long-range temporal dependencies in time-evolving physical systems, leading to suboptimal performance. This is especially challenging when dealing with sparse and unevenly distributed sensor data, where inferring the underlying dynamics requires understanding both spatial correlations and temporal evolution over extended periods.

Methods

The core innovation is the FR-Mamba architecture, a hybrid neural network that integrates two key components:

  • State Space Model (SSM) Branch: This branch leverages Mamba, a recent SSM architecture known for its efficiency in modeling long sequences with linear time complexity. It's designed to capture historical temporal features and state representations from sequential sensor data. The SSM branch uses a "selective input construction" mechanism to emphasize important time steps and a discretization process to adapt continuous-time SSMs for discrete data.
  • Fourier Neural Operator (FNO) Branch: This branch uses a 1D Fourier Neural Operator (FNO1d) to map sensor data into the frequency domain. This allows it to capture global spatial structures and non-local spatial features by leveraging frequency-domain transformations. The FNO1d is crucial for addressing the spatial modeling limitations of pure SSMs when sensor data is flattened into a sequence.

These two branches are combined within a dual-branch FNO-Mamba module. Features from both branches are fused using a gating mechanism. The overall architecture also includes a 2D Fourier Neural Operator (FNO2d) in the output domain to further enhance global spatial consistency and refine the reconstruction. The model is trained by minimizing the Mean Absolute Error (MAE) between predicted and ground-truth fields.

Datasets

The authors evaluate their method on a two-dimensional unsteady flow dataset simulating flow around a circular cylinder at a Reynolds number of Re = 100. This dataset consists of:

  • Observation data from 16 sensors uniformly distributed in the flow field.
  • 5000 temporally consecutive transient vortex field snapshots.
  • The resolution of each snapshot is (192, 112).
  • The first 4000 snapshots are used for training, and the last 1000 for validation/testing.

This dataset is described as being the same as used in previous works [24, 30].

Key Insights and Ideas

  • SSMs for Physical Fields: This paper pioneers the application of State Space Models (SSMs) to physical field reconstruction, demonstrating their effectiveness in capturing long-range temporal dependencies crucial for dynamic systems.
  • Synergy of FNO and SSM: The hybrid FNO-Mamba architecture effectively combines the strengths of SSMs for temporal modeling and FNOs for spatial feature extraction. This dual-branch approach is key to overcoming the limitations of each individual method.
  • Addressing Spatial Flattening: By using FNO1d, the model can implicitly learn spatial correlations even when sensor data is treated as a 1D sequence, a common issue in time-series processing.
  • Selective Temporal Encoding: The "selective input construction" for the SSM branch allows the model to dynamically focus on more informative time steps, improving efficiency and accuracy.
  • Hierarchical Representation: Stacking multiple FNO-Mamba blocks enables hierarchical representation learning, allowing the model to capture increasingly complex spatiotemporal dynamics.

Key Results

  • State-of-the-Art Accuracy: FR-Mamba significantly outperforms existing methods in flow field reconstruction tasks, achieving the lowest Mean Absolute Error (MAE) and Maximum Absolute Error (Max-AE) on the cylinder flow dataset.
  • Long Sequence Performance: The model demonstrates stable accuracy over long sequences, outperforming methods that struggle with long-term dependencies.
  • Robustness Across Regions: FR-Mamba shows consistent performance across different regions of the flow field (near-wall, wake, far-field), including areas with sharp vorticity gradients, where other methods falter.
  • Quantitative Improvements:
    • MAE on the cylinder flow dataset ranges from 0.30 to 0.32.
  • Max-AE is around 0.10.
  • Qualitative comparisons show FR-Mamba significantly reduces reconstruction error maps, with maximum errors dropping from values like 1.170 (Gappy MLP) to 1.02 × 10⁻³.
  • Ablation Studies:
    • Both FNO1d and FNO2d contribute significantly to performance. Removing them leads to a substantial increase in MAE (0.397) and Max-AE (0.355).
  • A moderate network depth (Nlayer = 10) with a hidden dimension of D = 8 yields the best trade-off between accuracy and efficiency.

Limitations and Future Work

  • Limitations:
    • While FR-Mamba shows strong performance, the paper notes that existing methods struggle with complex, high-dimensional spatiotemporal dynamics, particularly in regions with sharp vorticity gradients. This implies that while FR-Mamba is superior, there's still room for improvement in extremely complex scenarios.
  • The ablation study on network depth suggests that excessive depth can lead to parameter redundancy and overfitting, especially with limited training data.
  • Future Work:
    • The authors plan to explore further optimizations and adapt the model to more complex physical phenomena.
  • Specific future directions include extending the model to multi-phase flows and higher-dimensional physical fields.
  • The goal is to broaden its applicability to a wider range of scientific and engineering challenges.

Summary updated

Abstract

Physical field reconstruction (PFR) aims to predict the state distribution of physical quantities (e.g., velocity, pressure, and temperature) based on limited sensor measurements. It plays a critical role in domains such as fluid dynamics and thermodynamics. However, existing deep learning methods often fail to capture long-range temporal dependencies, resulting in suboptimal performance on time-evolving physical systems. To address this, we propose FR-Mamba, a novel spatiotemporal flow field reconstruction framework based on state space modeling. Specifically, we design a hybrid neural network architecture that combines Fourier Neural Operator (FNO) and State Space Model (SSM) to capture both global spatial features and long-range temporal dependencies. We adopt Mamba, a recently proposed efficient SSM architecture, to model long-range temporal dependencies with linear time complexity. In parallel, the FNO is employed to capture non-local spatial features by leveraging frequency-domain transformations. The spatiotemporal representations extracted by these two components are then fused to reconstruct the full-field distribution of the physical system. Extensive experiments demonstrate that our approach significantly outperforms existing PFR methods in flow field reconstruction tasks, achieving high-accuracy performance on long sequences.
BibTeX
@misc{long2025fr,
  author = {Jiahuan Long and Wenzhe Zhang and Ning Wang and Tingsong Jiang and Wen Yao},
  title = {{FR-Mamba: Time-Series Physical Field Reconstruction Based on State Space Model}},
  year = {2025},
  eprint = {2505.16083},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.16083},
}
Figure