Introduction
In large-scale scientific simulations—ranging from numerical weather prediction (NWP) to combustion modeling—the ability to accurately transfer data between different computational grids is essential. Whether moving from a non-uniform "dynamics" grid to a uniform "physics" grid, the interpolation method used must maintain the integrity of the underlying physical quantities. Standard high-order interpolation methods, while providing high accuracy, are notoriously susceptible to the Gibbs phenomenon. This phenomenon manifests as unphysical oscillations or "wiggles" near sharp gradients, which can lead to catastrophic errors such as negative values for density or moisture—quantities that are physically required to be positive.
This paper introduces a novel high-order interpolation framework that addresses these challenges by combining Essentially Non-Oscillatory (ENO) stencils with a new recursive bounding strategy for Newton polynomials. The resulting algorithm ensures physical properties like positivity and data boundedness are preserved without sacrificing the ability to capture sharp peaks and steep gradients.
The Challenge: Accuracy vs. Physicality
Existing solutions to the Gibbs phenomenon often involve a trade-off. Methods like Piecewise Cubic Hermite Interpolating Polynomials (PCHIP) are property-preserving but are generally limited to low-order (cubic) accuracy. Furthermore, these methods often "clip" or flatten out local extrema that fall between grid points, leading to a loss of information in regions of high physical interest.
The authors seek to overcome these limitations by developing a method that is both high-order and strictly constrained to physical bounds, allowing for the accurate representation of "hidden" peaks that lower-order methods might miss.
Methodology
The proposed approach is built upon the foundation of Newton polynomials and ENO schemes. The methodology consists of three primary components:
1. Adaptive Stencil Selection
Following the standard ENO philosophy, the algorithm evaluates neighboring data points to select the "smoothest" possible stencil for polynomial construction. This adaptive selection helps minimize oscillations from the outset by avoiding stencils that cross near-discontinuities.
2. Divided Difference Control
The core innovation of this work is the introduction of a control factor, $\lambda_j$, representing the ratio of successive divided differences. The authors mathematically derive recursive bounds ($B_j^-$ and $B_j^+$) for these ratios. By limiting these ratios, the growth of the polynomial is strictly controlled, preventing the wild excursions typical of high-order methods.
3. DBI and PPI Variants
The framework offers two distinct modes of operation:
- Data-Bounded Interpolation (DBI): This variant forces the interpolant to remain strictly within the range of the local data points. It is highly stable but can be overly restrictive.
- Constrained Positivity-Preserving Interpolation (PPI): This "relaxed" version ensures the value never drops below a user-defined minimum (e.g., zero) but allows the polynomial to exceed local data values by a small, controlled parameter ($\epsilon$). This flexibility allows the method to capture physical peaks that lie between grid points.
Experimental Validation
The authors benchmarked the method using several classic mathematical challenges designed to test interpolation stability and accuracy:
- The Runge Function: A standard test for high-order methods, where polynomials typically oscillate wildly near the edges of an interval.
- Smoothed Heaviside Function: Used to simulate steep gradients or near-discontinuities, testing the method's ability to handle sharp transitions without overshoot.
- LGL (Legendre-Gauss-Lobatto) Meshes: Non-uniform grids common in spectral element methods, testing the algorithm's robustness across different grid geometries.
Key Results and Insights
The study yielded several significant findings regarding the performance of the PPI and DBI methods:
- Superior Accuracy: In tests involving the Runge function, the PPI method achieved significantly lower $L^2$-error norms and higher convergence rates compared to the industry-standard PCHIP.
- Capturing Hidden Extrema: One of the most notable successes was the method's ability to recover peaks. While DBI and PCHIP tended to "flat-top" extrema that occurred between grid points, the PPI method (with a small $\epsilon$) successfully captured these peaks, providing a more faithful representation of the underlying function.
- Stability and Scalability: The recursive bounding math is applicable to any polynomial degree ($n$), allowing for 8th or 16th-order interpolation while maintaining guarantees against negative values. The method proved stable across both uniform and non-uniform grids, suggesting high utility for complex systems like the Navy’s NEPTUNE weather model.
Limitations and Future Directions
While the method represents a significant advancement, the authors identify areas for further refinement. The recursive bounding and adaptive stencil selection introduce more computational complexity than simple linear methods, necessitating future work in performance optimization and vectorization for modern hardware. Additionally, the PPI method requires a user-defined parameter ($\epsilon$) to control the allowable exceedance of data bounds; automating the selection of this parameter for different physical variables remains an open research question. Finally, while 2D results were demonstrated via tensor products, extending the method to full 3D production environments is a priority for future development.
Summary updated