Scientific data interpolation with low dimensional manifold model.

Wei Zhu, Bao Wang, Richard C. Barnard, Cory D. Hauck, Frank Jenko, Stanley J. Osher

TL;DR

This paper introduces a novel method for interpolating scientific data, even when large portions are missing, by leveraging the idea that data often lies on a low-dimensional "manifold." Their approach, the Low Dimensional Manifold Model (LDMM), shows superior performance compared to existing methods for data reconstruction and offers a promising, albeit less accurate, alternative for data compression.

What Problem Does This Paper Address?

Scientific data, whether from simulations or experiments, is often incomplete due to various constraints like limited sensors, computational costs, or intentional sampling for compression. Reconstructing this missing information accurately is crucial for analysis and understanding. Traditional methods often rely on specific physical models or struggle when a significant amount of data (e.g., over 95%) is missing, especially with irregular sampling patterns. This paper aims to address this challenge by proposing a general-purpose interpolation technique that can handle both regular and irregular sparse data.

Methods

The core of the proposed method is the Low Dimensional Manifold Model (LDMM). The key insight is that patches of data (small local regions) from scientific datasets often lie on a lower-dimensional manifold than the ambient space they inhabit. This "patch manifold" property is used as a regularizer in a variational formulation. The problem is then solved using an alternating minimization approach: first, the data is updated to better fit the observed samples and the current manifold estimate, and second, the manifold is updated to better represent the data patches.

The paper details the discretization of the resulting Euler-Lagrange equation using the weighted graph Laplacian. This is a crucial step for practical implementation, allowing the algorithm to efficiently handle the complex manifold structure. The method is designed to work with both 2D and 3D data.

Datasets

The authors evaluate their LDMM approach on a variety of scientific datasets, including:

  • 3D plasma (magnetic field): From gyrokinetic simulations of Alfvénic turbulence.
  • 3D/2D lattice: A benchmark problem for angular discretization methods in radiation transport.
  • 3D/2D plasma (distribution function): Also from gyrokinetic simulations.
  • 2D vortex: From a numerical solution of the Orszag-Tang vortex system.

These datasets represent diverse physical phenomena and data structures, allowing for a comprehensive evaluation of LDMM's generalization capabilities. The paper also discusses different sampling strategies, including random subsampling (5% and 10%) and regular downsampling (factors of 2 and 4).

Key Insights and Ideas

  • The "Patch Manifold" Hypothesis: The fundamental idea that local data patches in scientific datasets often reside on a lower-dimensional manifold is a powerful concept. This intrinsic dimensionality acts as a strong prior for reconstruction.
  • Generalizability Beyond Images: While LDMM was initially developed for image processing, this work demonstrates its applicability and effectiveness for a broader range of scientific data, highlighting the universality of the low-dimensional manifold assumption.
  • Weighted Graph Laplacian for Efficient Manifold Handling: Discretizing the manifold operations using the weighted graph Laplacian is key to making the LDMM computationally feasible and efficient, especially compared to methods relying on more complex integral formulations.
  • LDMM as a Dual-Purpose Tool: The paper explores LDMM not only for interpolation but also for data compression, revealing trade-offs between accuracy and compression efficiency.

Key Results

  • Superior Interpolation Performance: LDMM consistently outperforms other state-of-the-art methods (like Exemplar-Based Interpolation (EBI) and Piecewise Linear Estimator (PLE)) in terms of accuracy (measured by L1, L2, L∞ norms, and PSNR) for both regular and irregular sampling, especially when a large amount of data is missing.
  • Robustness to Sampling: The method shows strong performance even with very sparse sampling (as low as 5%), where other methods often fail.
  • Data Compression Trade-offs: As a data compression technique, LDMM is easier to implement and faster for reconstructing partial data compared to traditional methods. However, it generally yields lower accuracy than methods like DCT, SVD, or Tucker decomposition when aiming for the same compression rate. Despite this, it shows competitive results on some complex datasets.
  • Fast Convergence: Numerical results indicate that LDMM converges relatively quickly, typically within 10 iterations.

Limitations and Future Work

  • Data Compression Accuracy: The authors acknowledge that LDMM, in its current form, does not achieve the same level of accuracy as established compression methods that have access to the full dataset.
  • Future Work on Compression: The paper suggests significant room for improvement in LDMM's data compression capabilities. Future work will focus on:
  • Strategic Sampling: Instead of random sampling, strategically choosing pixels based on prior information.
  • Modifying the Algorithm: Incorporating gradient information or specific entries from weight matrices into the sampling process.
  • Optimizing for Compression: Further developing LDMM to better balance accuracy and compression efficiency.

Summary updated

Abstract

Highlights • A low dimensional manifold model is applied to the interpolation of scientific data from sparse sampling. • The low dimensionality of the patch manifold for general scientific data sets is used as a regularizer. • The proposed algorithm consistently outperforms other state-of-the-art methods in this field. • The performance of the proposed method as a data compression technique is compared to other standard methods. Abstract We propose to apply a low dimensional manifold model to scientific data interpolation from regular and irregular samplings with a significant amount of missing information. The low dimensionality of the patch manifold for general scientific data sets has been used as a regularizer in a variational formulation. The problem is solved via alternating minimization with respect to the manifold and the data set, and the Laplace–Beltrami operator in the Euler–Lagrange equation is discretized using the weighted graph Laplacian. Various scientific data sets from different fields of study are used to illustrate the performance of the proposed algorithm on data compression and interpolation from both regular and irregular samplings. Previous article in issue Next article in issue
BibTeX
@article{zhu2018scientific,
  author = {Wei Zhu and Bao Wang and Richard C. Barnard and Cory D. Hauck and Frank Jenko and Stanley J. Osher},
  title = {{Scientific data interpolation with low dimensional manifold model.}},
  journal = {Journal of Computational Physics},
  year = {2018},
  doi = {10.1016/j.jcp.2017.09.048},
}
Figure