Novel distance-based masking and adaptive alpha-shape methods for CNN-ready reconstruction of arbitrary 2D CFD flow domains

This paper introduces a novel reconstruction framework featuring distance-based masking and adaptive alpha-shape methods to accurately recover physical boundaries from scattered 2D CFD data for CNN-ready grids, offering significant speedups, minimal tuning requirements, and a companion web application for end-to-end processing.

The Gist

The Big Picture: Turning a "Cloud" into a "Picture"

Imagine you are a chef trying to teach a robot (a CNN, or Convolutional Neural Network) how to cook a specific dish. The robot is very smart, but it only understands recipes written on a perfect, uniform grid of squares (like a chessboard).

However, the data you have from your experiments (the CFD flow data) is messy. It's like a cloud of floating dust particles scattered randomly in the air. Some areas are dense with dust, others are sparse. If you try to force this scattered cloud onto the robot's perfect chessboard grid, the robot gets confused. It tries to connect the dots, but instead of seeing the true shape of the dish, it draws a giant, smooth, convex bubble around everything (like wrapping the dust in a balloon). This "balloon" includes empty space where no dust exists, which would make the robot learn the wrong physics.

The Goal of this Paper:
The authors built a "smart filter" that takes that messy cloud of data and cuts away the empty space, leaving only the true, complex shape of the dish. They want to give the robot a clean, accurate map so it can learn effectively.

---

The Three Methods: How to Cut the Shape

The paper tests three different ways to cut out the true shape from the messy data.

1. The "Distance Ruler" Method (Distance-Based Masking)

• The Analogy: Imagine you have a bucket of marbles (your data points) scattered on a table. You want to know which spots on the table are "safe" (inside the shape) and which are "unsafe" (outside).
• How it works: You take a ruler and measure the distance from every spot on the table to the nearest marble.
  • If a spot is close to a marble (within a set distance, like 1 inch), you mark it as "Inside."
  • If a spot is far away, you mark it as "Outside."
• Why it's great: It's incredibly fast and simple. You don't need to know the complex rules of geometry; you just need a ruler. The authors found that setting the ruler to the size of the grid squares themselves works perfectly for almost every shape they tried.
• Speed: It's like a lightning bolt—hundreds of times faster than the other methods.

2. The "Rubber Band" Method (Classical $\alpha$-Shape)

• The Analogy: Imagine you have a giant, stretchy rubber band. You throw it over your scattered marbles.
  • If the rubber band is loose (large setting), it snaps tight around the outermost marbles, creating a smooth, round shape (the "convex hull"). It misses all the dents and curves.
  • If you tighten the rubber band (small setting), it sinks into the gaps between marbles, capturing the true, bumpy shape.
• The Problem: The "tightness" setting (called $\alpha$) is tricky. If you set it too loose, you get a blob. If you set it too tight, the rubber band snaps and breaks the shape into tiny, disconnected pieces. You have to manually tweak this setting for every single new shape you try.
• Speed: It's slow, like trying to tie a complex knot by hand.

3. The "Self-Adjusting Rubber Band" (Adaptive $\alpha$-Shape)

• The Analogy: This is the same rubber band, but now it's made of "smart material." It can feel how close the marbles are to each other.
  • In crowded areas (dense marbles), it tightens up to capture fine details.
  • In sparse areas (few marbles), it loosens up so it doesn't break.
• Why it's great: You don't have to manually tweak the settings as much. It adapts to the data automatically.
• Speed: It's faster than the old rubber band method, but still slower than the simple "Distance Ruler."

---

The "Inflation" Trick (Boundary Inflation)

Even with the best methods, sometimes the edge of the shape is so thin that the robot's grid squares miss a few marbles right on the border. It's like trying to fit a square peg into a round hole; the corner might just barely miss.

• The Fix: The authors added a tiny "inflation" step. After cutting the shape, they blow it up by a microscopic amount (like 0.2%).
• The Result: This tiny puff ensures no marbles are left outside the door, but it doesn't puff the shape out so much that it includes empty space. It's like putting a tiny, invisible coat of paint on the edge to make sure everything is covered.

---

The Verdict: Which Method Wins?

The authors tested these methods on four different complex shapes (like a turbine blade, a Y-shaped pipe, and a nozzle). Here is what they found:

1. The Winner (Distance-Based): This is the recommended default. It is the "Swiss Army Knife" of the group. It is:
   • Super Fast: 500 to 800 times faster than the old method.
   • Stable: You set the ruler once, and it works for everything.
   • Accurate: It captures the shape perfectly.
2. The Runner-Up (Adaptive $\alpha$-Shape): This is a great backup plan. If you don't know the size of the grid squares (so you can't use the ruler), this method is your best friend. It's smart and adaptable, though a bit slower.
3. The Old Guard (Classical $\alpha$-Shape): This is the "fussy" one. It works well if you are an expert who knows exactly how to tune the settings for every single new shape, but it's too slow and difficult for general use.

The "Magic Tool" (Web Application)

To make this useful for everyone, the authors didn't just write a paper; they built a free website.

• You can upload your messy data files.
• You can play with the settings (like the ruler length or the rubber band tightness).
• The website instantly draws the perfect shape and gives you the clean data ready for your AI robot.

Summary in One Sentence

This paper teaches us how to turn messy, scattered scientific data into a clean, perfect picture for AI to learn from, using a super-fast "distance ruler" method that is easy to use and incredibly accurate.

Technical Summary

1. Problem Statement

Convolutional Neural Networks (CNNs) require structured, uniform grid data to function effectively. However, Computational Fluid Dynamics (CFD) simulations typically produce scattered, unstructured datasets (e.g., from irregular meshes or point clouds).

• The Core Issue: Directly interpolating scattered CFD data onto a uniform Cartesian grid creates a "convex-hull" envelope. This distorts the true geometry by activating non-physical regions (spurious voxels) in concave areas, gaps, or disconnected components.
• The Consequence: Training CNNs on these distorted fields leads to physical inconsistencies, as the network learns from data in regions where no fluid actually exists.
• The Goal: Develop a robust reconstruction framework to recover accurate, concave physical boundaries from scattered data and generate "CNN-ready" binary masks that exclude non-physical regions while preserving geometric fidelity.

2. Methodology

The authors propose a three-stage pipeline to reconstruct the domain from scattered CFD data ($\mathcal{P}$) onto a structured grid ($\mathcal{G}$):

A. Grid Construction and Interpolation

• An axis-aligned bounding box is defined around the scattered data.
• A uniform Cartesian grid is constructed.
• Scattered data is interpolated onto this grid using piecewise-linear interpolation based on Delaunay tessellation (barycentric coordinates). This initial step inevitably creates the convex-hull distortion that needs correction.

B. Boundary Recovery Strategies

The paper introduces and evaluates three distinct methods to generate the binary mask (identifying physical vs. non-physical voxels):

1. Distance-Based Masking (Novel):

   • Mechanism: Computes the Euclidean distance from every grid node to the nearest scattered sample point.
   • Thresholding: A binary mask is created where nodes within a distance threshold $\tau$ are classified as "inside."
   • Refinement: A morphological closing operation (dilation followed by erosion) is applied to fill small holes and smooth the boundary.
   • Key Feature: The threshold $\tau$ is set to the minimum CFD grid spacing ($\Delta$), making it robust and parameter-free across different geometries.

2. Classical $\alpha$-Shape Reconstruction (Baseline):

   • Mechanism: Uses the Delaunay triangulation of the point set. It filters simplices based on their circumradius ($r_\sigma$). Only simplices with $r_\sigma \leq \alpha$ are retained.
   • Limitation: Requires a global $\alpha$ parameter. A single $\alpha$ often fails to handle varying sampling densities or complex concavities (too small fragments the shape; too large over-smooths it).

3. Adaptive $\alpha$-Shape Reconstruction (Novel):

   • Mechanism: Normalizes the $\alpha$ parameter using local data resolution. Instead of a fixed length, it uses a dimensionless control parameter $\beta$ scaled by the mean Delaunay edge length ($\bar{e}$) of the dataset: $\alpha_a = \beta / \bar{e}$.
   • Advantage: This makes the method resolution-independent and robust to varying sampling densities without requiring geometry-specific tuning.

C. Post-Processing: Boundary Inflation

• To address "edge misses" caused by sub-voxel misalignment (where valid boundary samples fall just outside the mask), a minimal boundary inflation (dilation by a factor $\eta \approx 0.2\%$) is applied. This recovers missed samples with negligible activation of non-physical regions.

D. Evaluation Metrics

A new suite of topology-aware metrics was introduced to quantitatively assess the methods:

• Point Recall (PR): Fraction of original samples retained.
• Ghost Fraction (GF): Fraction of active voxels far from data (non-physical regions).
• Active Volume Fraction (AVF): Relative size of the mask.
• Connected Components (CC): Measures topological fragmentation.
• Intersection over Union (IoU), Precision, Recall: Comparisons against a reference mask.

3. Key Contributions

1. Novel Distance-Based Method: A highly efficient, robust masking technique that uses a simple distance threshold derived from grid spacing. It eliminates the need for complex boundary definitions or normal estimation.
2. Adaptive $\alpha$-Shape Formulation: A new formulation that normalizes the $\alpha$ parameter by local edge length, removing the need for manual, geometry-specific tuning of the scale parameter.
3. Comprehensive Metric Suite: Introduction of a quantitative framework (PR, GF, IoU, etc.) to objectively evaluate domain reconstruction, moving beyond visual inspection.
4. Boundary Inflation Strategy: A lightweight post-processing step that significantly improves data retention (up to ~3%) with minimal cost to physical consistency.
5. Operational Web Tool: Development of a companion web application (CFDtoCNN.streamlit.app) that allows users to upload 2D ASCII datasets, tune parameters, and export CNN-ready masks.

4. Results and Performance

The methods were tested on four complex 2D internal flow geometries: a sudden expansion-contraction duct, a Y-shaped bifurcation, a converging-diverging nozzle, and a curved turbine passage.

• Accuracy & Robustness:

  • Distance-Based: Achieved near-perfect retention (PR $\approx$ 1) and near-zero ghost activation (GF $\approx$ 0) with a single universal threshold ($\tau = \Delta$). It was the most stable across all geometries.
  • Adaptive $\alpha$-Shape: Performed comparably to the distance-based method when $\beta=1$, maintaining stability without geometry-specific tuning.
  • Classical $\alpha$-Shape: Required significant manual tuning of $\alpha$ for each geometry (e.g., $\alpha=10$ for nozzles vs. $\alpha=1000$ for bifurcations) to avoid fragmentation or over-smoothing.

• Computational Efficiency:

  • Distance-Based: Extremely fast, generating masks in 15–18 ms. It is 500–800$\times$ faster than the classical $\alpha$-shape.
  • Adaptive $\alpha$-Shape: Significantly faster than the classical variant (1.7–2.6$\times$ speedup), taking seconds rather than tens of seconds.
  • Classical $\alpha$-Shape: The slowest method, requiring seconds per mask due to complex geometric calculations.

• Impact of Boundary Inflation:

  • Applying the minimal inflation step increased Point Recall by 0.18% to 2.96% across methods while increasing Ghost Fraction by less than 0.08%, effectively solving the edge-miss issue.

5. Significance and Conclusion

This work addresses a critical bottleneck in applying Deep Learning to CFD: the preprocessing of unstructured data into structured formats.

• Practical Impact: The proposed framework enables the creation of high-fidelity, physically consistent datasets for training CNNs, which is essential for developing data-driven surrogates for fluid dynamics.
• Recommendation: The authors recommend the Distance-Based Method as the default choice due to its superior speed, accuracy, and lack of parameter tuning. The Adaptive $\alpha$-Shape is recommended as a strong alternative when grid-spacing information is unavailable.
• Accessibility: By providing a web-based tool, the authors democratize access to these advanced reconstruction techniques, facilitating reproducible research and efficient data preparation for machine learning in engineering.