Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WindMiL: Equivariant Graph Learning for Wind Loading Prediction

Themistoklis Vargiemezis, Charilaos Kanatsoulis, Catherine Gorlé

TL;DR

WindMiL addresses the need for scalable wind loading surrogates by coupling a large LES-derived, geometry-interpolated dataset with a reflection-equivariant GNN. The method enforces horizontal-plane symmetry by averaging predictions over original and reflected inputs, achieving high accuracy for both mean and variability of the surface pressure coefficient $C_p$ in both interpolation and extrapolation, and maintaining performance under symmetry transformation where non-equivariant baselines falter. The dataset uses Signed Distance Function interpolation across three basis roof geometries, yielding 462 cases plus rotations, while LES computations provide the high-fidelity targets. Practically, WindMiL enables efficient, physically consistent wind-load predictions for design exploration at a fraction of the LES cost, with clear avenues for extension to urban-scale and multi-condition scenarios.

Abstract

Accurate prediction of wind loading on buildings is crucial for structural safety and sustainable design, yet conventional approaches such as wind tunnel testing and large-eddy simulation (LES) are prohibitively expensive for large-scale exploration. Each LES case typically requires at least 24 hours of computation, making comprehensive parametric studies infeasible. We introduce WindMiL, a new machine learning framework that combines systematic dataset generation with symmetry-aware graph neural networks (GNNs). First, we introduce a large-scale dataset of wind loads on low-rise buildings by applying signed distance function interpolation to roof geometries and simulating 462 cases with LES across varying shapes and wind directions. Second, we develop a reflection-equivariant GNN that guarantees physically consistent predictions under mirrored geometries. Across interpolation and extrapolation evaluations, WindMiL achieves high accuracy for both the mean and the standard deviation of surface pressure coefficients (e.g., RMSE $\leq 0.02$ for mean $C_p$) and remains accurate under reflected-test evaluation, maintaining hit rates above $96\%$ where the non-equivariant baseline model drops by more than $10\%$. By pairing a systematic dataset with an equivariant surrogate, WindMiL enables efficient, scalable, and accurate predictions of wind loads on buildings.

WindMiL: Equivariant Graph Learning for Wind Loading Prediction

TL;DR

WindMiL addresses the need for scalable wind loading surrogates by coupling a large LES-derived, geometry-interpolated dataset with a reflection-equivariant GNN. The method enforces horizontal-plane symmetry by averaging predictions over original and reflected inputs, achieving high accuracy for both mean and variability of the surface pressure coefficient in both interpolation and extrapolation, and maintaining performance under symmetry transformation where non-equivariant baselines falter. The dataset uses Signed Distance Function interpolation across three basis roof geometries, yielding 462 cases plus rotations, while LES computations provide the high-fidelity targets. Practically, WindMiL enables efficient, physically consistent wind-load predictions for design exploration at a fraction of the LES cost, with clear avenues for extension to urban-scale and multi-condition scenarios.

Abstract

Accurate prediction of wind loading on buildings is crucial for structural safety and sustainable design, yet conventional approaches such as wind tunnel testing and large-eddy simulation (LES) are prohibitively expensive for large-scale exploration. Each LES case typically requires at least 24 hours of computation, making comprehensive parametric studies infeasible. We introduce WindMiL, a new machine learning framework that combines systematic dataset generation with symmetry-aware graph neural networks (GNNs). First, we introduce a large-scale dataset of wind loads on low-rise buildings by applying signed distance function interpolation to roof geometries and simulating 462 cases with LES across varying shapes and wind directions. Second, we develop a reflection-equivariant GNN that guarantees physically consistent predictions under mirrored geometries. Across interpolation and extrapolation evaluations, WindMiL achieves high accuracy for both the mean and the standard deviation of surface pressure coefficients (e.g., RMSE for mean ) and remains accurate under reflected-test evaluation, maintaining hit rates above where the non-equivariant baseline model drops by more than . By pairing a systematic dataset with an equivariant surrogate, WindMiL enables efficient, scalable, and accurate predictions of wind loads on buildings.

Paper Structure

This paper contains 10 sections, 5 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Dataset generation
  3. Geometry parameterization
  4. Large-eddy Simulations
  5. Model architecture
  6. Experiments
  7. Interpolation performance
  8. Extrapolation performance
  9. Qualitative Analysis
  10. Conclusions

Figures (7)

  • Figure 1: Contour plots of mean $C_p$ at $-45^{\circ}$ and $-45^{\circ}$ wind incidence.
  • Figure 2: Convex hull of the three basis geometries. Points inside the convex hull correspond to interpolated buildings.
  • Figure 3: SDF interpolation between a sphere and a cube, per Eq. (1) for different $\alpha$.
  • Figure 4: View of the mesh used for the large-eddy simulations, with side views of refinement zones shown: (left) near the building and (middle) in the surrounding area. (right) Top view of instantaneous velocity magnitude contours showing the turbulent flow around the building.
  • Figure 5: Visualization of dataset splits in the barycentric shape space. The convex hull is spanned by the three basis roof geometries (flat, gable, hip), and each point corresponds to one interpolated building.
  • ...and 2 more figures