This repository contains the Sound Transmission Loss (STL) datasets and Machine-Learning models described in the paper entitled "On Machine Learning-Driven Surrogates for Sound Transmission Loss Simulations".
Six physics-driven models of STL were used to create the datasets:
- analytical STL solution for infinite plate (Infinite_Isotropic folder),
- finite plate STL approximated with a correction factor based on Rayleigh-integral (Correction_Factor folder),
- analytical Modal Summation (MS) approach (Modal_Summation folder),
- MS approach with one-third octave band average (Modal_Summation_avg folder),
- numerical solution using Finite Element Method (FEM) modeled in COMSOL Multiphysics® (Comsol folder),
- FEM solution with one-third octave band average (Comsol_avg folder).
Four Machine Learning (ML) algorithms are used to create the surrogates:
- Neural Networks (NN),
- Guassian Process Regressor (GPR),
- Random Forest (RF),
- Gradient Boosting Trees (GBT).
The STL variables are the plate thickness h, density ρ, Young's Modulus E, Poisson's ratio ν, damping factor η, and, for the case of finite plates, the plate width a and length b. The codes also permit to include the mass density m, the bending stiffness D_R and the resonance coefficient term R as physics-guided inputs to the ML models. The ML-driven surrogate outputs are the STL values for each frequency in the df_freq files. In addition, Mean Decrease in Impurity (MDI)-based Sensitivity Analysis is performed with RF constructed with Multi-Output Regressor.
If you use these datasets and models in your research, please cite:
@article{Cunha2022,
author = {Barbara Zaparoli Cunha and Abdel-Malek Zine and Mohamed Ichchou and Christophe Droz and Stéphane Foulard},
title = {On Machine Learning-Driven Surrogates for Sound Transmission Loss Simulations},
year = {2022}
}