Intelligent Automated Grid Generation for Numerical Simulations

Ke-Thia Yao, Andrew GelseyAAAI 1994

Introduction

In Computational Fluid Dynamics (CFD), the primary objective is to solve complex Partial Differential Equations (PDEs) to predict fluid behavior. However, the most significant bottleneck in this process is not the solver itself, but "gridding"—the task of discretizing a physical object's surface into a mesh of panels. Historically, this has been a labor-intensive process requiring human experts to manually partition surfaces into "griddable" patches. This manual intervention is time-consuming, prone to error, and prevents the realization of fully automated engineering design systems.

This paper presents an "intelligent gridder" designed to automate this critical step. By integrating topological analysis with physics-based heuristics—specifically reasoning about fluid streamlines—the system can automatically partition and discretize complex geometries that previously demanded significant manual oversight.

Methodology: The Boundary Surface Representation (BSR)

The core of the proposed system is a new language called Boundary Surface Representation (BSR). Unlike standard geometric models that focus solely on shape, BSR explicitly encodes both geometry and topology (the connectivity between parts).

BSR treats surface mappings as "black boxes," which allows it to handle diverse mathematical definitions—ranging from simple algebraic formulae to complex B-splines—within a uniform framework. This abstraction is crucial for managing the complex intersections found in real-world engineering models.

Streamline-Based Reasoning

The "intelligence" of the gridder stems from its ability to reason about the physics of the problem. The system approximates how fluid would flow over the object by projecting a free-stream vector onto the surface to generate streamlines. These streamlines serve as the primary heuristic for two critical tasks:

  1. Automated Partitioning: To satisfy the requirements of numerical solvers like PMARC, surfaces must be "simply connected" (i.e., they cannot have holes). The system uses streamlines to "cut" holes in surfaces—for instance, where a keel meets a hull—ensuring the resulting patches are topologically valid for discretization.
  2. Reparametrization and Distribution: The system groups boundary edges into four-sided regions suitable for transfinite interpolation. Streamlines help decide which edges are "parallel" or "orthogonal" to the flow. Furthermore, the system automatically clusters grid lines in regions of high physical interest, such as near "sources" and "sinks" where flow velocity changes rapidly.
Figure 6: Streamline-based reasoning and automated "cuts" used to partition the hull, keel, and winglet into simply-connected regions.

Evaluation: The Stars & Stripes Benchmark

The system was evaluated using a complex, real-world engineering challenge: the geometry of the Stars & Stripes racing yacht, the 1987 America’s Cup winner. This model, consisting of an ellipsoid hull, a keel, and a winglet, represents a significant topological hurdle. The intersections between these components create complex boundaries that are notoriously difficult to automate.

Key Results and Insights

The implementation successfully generated valid grids for the yacht geometry without human intervention. This achievement was notable as it surpassed the capabilities of contemporary automated gridders which typically required pre-partitioned patches.

Figure 1: The final automatically generated grid for the Stars & Stripes yacht, showing the integrated hull, keel, and winglet mesh.

Physics-Informed Geometry

A central insight of this work is that gridding should not merely follow the geometric contours of an object; it should follow the logic of the fluid. By aligning the grid with streamlines, the system produces meshes that are inherently optimized for the underlying numerical solver, leading to better accuracy and stability.

Numerical Convergence

Beyond qualitative success, the authors conducted a convergence study. They demonstrated that as the density of the automatically generated grid increased, key engineering metrics—such as the "effective draft" of the keel—converged to stable values. This confirms that the automated process produces numerically sound grids capable of generating reliable engineering data.

Limitations and Future Directions

While the current system marks a significant advancement, it has limitations. The reasoning is currently tailored to potential flows (Laplace’s equation) and may require adjustment for more turbulent or non-linear regimes. Additionally, the heuristic for grouping boundary arcs, while effective in tests, could potentially struggle with extremely convoluted geometries.

Future work aims to close the loop by incorporating feedback from preliminary simulations. By analyzing initial pressure coefficients, the system could perform local grid refinement in areas of high error. The authors also envision extending this "feature-based" reasoning to other physical domains, such as using temperature profiles to guide grid generation in heat transfer simulations.

Summary updated

BibTeX
@inproceedings{yao1994intelligent,
  author = {Ke-Thia Yao and Andrew Gelsey},
  title = {{Intelligent Automated Grid Generation for Numerical Simulations}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure