Abstract
Scientific data reconstruction is a critical challenge in fields ranging from plasma physics to nuclear engineering, where data acquisition is often limited by physical constraints or high costs. This paper introduces a novel interpolation framework based on the Low Dimensional Manifold Model (LDMM). By treating local data patches as points on a low-dimensional manifold, the proposed method utilizes a weighted graph Laplacian to reconstruct complex 2D and 3D datasets from as little as 5% of the original samples. Experimental results demonstrate that LDMM significantly outperforms traditional signal processing and physics-blind techniques, particularly in preserving oscillatory structures and sharp interfaces.
Introduction
Scientific datasets—such as seismic surveys, plasma simulations, and radiation transport models—are frequently incomplete. This sparsity arises from physical limitations (e.g., a finite number of sensors), economic constraints, or intentional downsampling for data compression. The central challenge is "interpolation": the accurate estimation of missing values from a sparse set of observations.
Existing interpolation methods generally fall into two categories. Physics-based models offer high accuracy but are often highly specialized to specific domains (e.g., fluid dynamics) and lack generalizability. Conversely, general-purpose signal processing models, such as wavelets or Total Variation (TV) minimization, often struggle when data is extremely sparse (missing >95%) or when sampling occurs on a regular grid rather than a random one. This paper addresses these gaps by providing a high-performance, general-purpose framework that exploits the intrinsic geometric structure of scientific data.
The Low Dimensional Manifold Model (LDMM)
The Patch Manifold Concept
The core innovation of this research is the application of the "patch manifold" concept to scientific data. Rather than processing the dataset as a single monolithic block, the LDMM algorithm decomposes the data into small, overlapping patches (e.g., $6 \times 6 \times 4$ cubes for 3D data).
When these patches are treated as points in a high-dimensional space, they do not distribute randomly. Instead, they tend to cluster around a low-dimensional manifold—much like a 2D sheet of paper crumpled within a 3D space. This observation suggests that scientific data possesses a high degree of local redundancy, meaning the intrinsic dimension of the patch manifold is significantly lower than the total number of voxels within a patch.
Algorithmic Implementation
The interpolation is formulated as a variational problem: find the dataset that minimizes the dimension of its patch manifold while remaining consistent with the known sample points. To solve this, the authors employ an Alternating Minimization strategy, switching between updating the data values and refining the manifold structure.
To ensure computational efficiency, the authors utilize a Weighted Graph Laplacian (WGL) to discretize the Laplace-Beltrami operator. This technical contribution allows the system to solve a single linear equation per iteration, providing a substantial speed advantage over previous point-integral methods and making the reconstruction of large 3D volumes feasible.
Experimental Evaluation
The LDMM was evaluated across six distinct scientific datasets characterized by complex oscillatory structures and sharp interfaces (shocks):
- 3D Plasma (Magnetic Field): Fluctuations from gyrokinetic simulations.
- 3D/2D Lattice: Radiation transport benchmarks for nuclear reactor assemblies.
- 3D/2D Plasma (Distribution Function): Ion species data in 5D phase space.
- 2D Vortex: Numerical solutions to the Orszag-Tang vortex system.
Key Results
LDMM was benchmarked against state-of-the-art methods, including Piecewise Linear Estimators (PLE) and Exemplar-Based Interpolation (EBI).
- Performance at Low Sampling Rates: At 5% random sampling, LDMM consistently achieved superior Peak Signal-to-Noise Ratios (PSNR). For the 2D Vortex dataset, LDMM reached 39.09 dB, significantly outperforming PLE, which reached only 24.84 dB.
- Regular Grid Robustness: Many patch-based methods suffer from aliasing artifacts when data is sampled on a regular grid. LDMM demonstrated resilience in these scenarios, improving the PSNR of 3D Lattice data from 30.01 dB (using Cubic Spline initialization) to 39.93 dB.
- Visual Fidelity: Qualitatively, LDMM reconstructions appear smoother and more physically plausible than competing methods, which often produce "patchy" artifacts or over-smoothed edges.
Discussion and Future Work
The success of LDMM highlights that scientific data is inherently redundant. By using the manifold's dimension as a regularizer, the model naturally preserves geometric structures like waves and edges better than simple smoothing filters. Furthermore, LDMM can function as a "sampling-based" compressor; by storing only a small fraction of random points, the algorithm can effectively "imagine" the missing data based on the learned manifold structure.
However, limitations remain. While excellent for interpolation, LDMM is currently less efficient than standard transform-based compression (like DCT or Tucker Decomposition) when the full dataset is available for encoding. Additionally, the construction of weight matrices remains computationally intensive, though parallelization offers a path toward mitigation. Future research will explore "strategic sampling"—choosing sample points based on data gradients or prior knowledge—to further enhance reconstruction accuracy.
Summary updated