JanvDelden/minimizing_structural_vibrations

Overview

This repository contains two main datasets: a primary dataset with 51,000 plate geometries (each with 15 frequency solutions) intended for training and validation, and a test dataset with 500 geometries (each with 300 frequency solutions).

Dataset files

• D50k_edited.h5: Dataset with 51,000 plate geometries and numerical solutions for 15 frequencies in [1, 300] Hz.
• D500_edit.h5: Dataset with 500 plate geometries and numerical solutions for 300 frequencies in [1, 300] Hz.
• moments_d50k.pt: Mean and standard deviation for the (squared and log-transformed) velocity fields and (untransformed) frequency response. The mean is computed per frequency in 1 Hz intervalls in [1, 300]. The standard deviation is computed as a scalar over all frequencies.
• D50k_complex_field.h5: Extended version of D50k_edited.h5, that includes the full complex velocity fields.

Data Format & Usage

The datasets are saved in the HDF5 file format (.h5) and compressed with the "Blosc" compression filter. To read these files, you will need the hdf5plugin Python package. Include the following in your Python script:

import h5py
import hdf5plugin

The HDF5 files contain the following keys:

• bead_patterns: The mesh geometry (unit: meters).
• z_vel_mean_sq: The frequency response (unit: decibels).
• z_vel_abs: The magnitude of velocity fields perpendicular to the plate (unit: meters per second).
• z_vel: Complex velocity fields (Only in D50k_complex_field.h5), replaces z_vel_abs (unit: meters per second).
• phy_para: An array of scalar physical properties in following order:
  • Rotational stiffness at the boundary in range [0, 100] (unit: Nm/rad).
  • Excitation point x-coordinate in range [0.2, 0.8] (unitless/normalized).
  • Excitation point y-coordinate in range [0.2, 0.8] (unitless/normalized).
• frequencies: The frequencies for frequency response and velocity fields (unit: hertz)

Model Weights

Alongside the data, we provide the weights of the trained models for design optimization. Please refer to the example notebook in the associated code repository for an example how to use the models.

• flow_matching_0331.ckpt: Flow matching model weights.
• regression_075_noise_0312.ckpt: Regression model weights, trained with noisy beading patterns.
• regression_no_noise_0526.ckpt: Regression model weights, trained with clean beading patterns.

More Information

Please refer to the paper 'Minimizing Structural Vibrations via Guided Flow Matching Design Optimization' for a full description of the dataset and mechanical model: https://arxiv.org/abs/2506.15263

This dataset has also been used in Predicting Eigenmode Decompositions in Vibroacoustic Systems.

Code for using this dataset is available here: https://github.com/ecker-lab/Optimizing_Vibrating_Plates