Learning acoustic responses from experiments: A multiscale-informed transfer learning approach.

Area of Science:

• Acoustics
• Materials Science
• Machine Learning

Background:

• Learning acoustical responses typically requires extensive experimental data.
• Data limitations pose challenges in developing accurate acoustic models.
• Predicting material acoustic properties is crucial for effective design.

Purpose of the Study:

• To develop a methodology for learning acoustical responses from limited experimental datasets.
• To enable accurate prediction of sound absorption coefficients using a novel approach.
• To facilitate efficient exploration of parameter spaces for acoustic materials.

Main Methods:

• A multiscale-informed encoder was used to create a finite-dimensional learning setting.
• A neural network model was trained using transfer learning and knowledge from a multiscale surrogate.
• The sound absorption coefficient was measured using a two-microphone method and predicted via a hybrid numerical approach (Johnson-Champoux-Allard-Pride-Lafarge model).

Main Results:

• The methodology successfully approximated the relationship between micro-/structural parameters and experimental acoustic response.
• Accurate predictions of sound absorption coefficients were achieved with a small training dataset (ten samples).
• The approach demonstrated effectiveness even with limited physical data.

Conclusions:

• The proposed methodology enables acoustic model identification and validation under data constraints.
• This approach facilitates efficient parameter space exploration for acoustic materials design.
• The study highlights the potential of machine learning in acoustics with limited data.