Physics-informed machine learning for micro-perforated panels: Reproducible prediction, uncertainty, and design...

Area of Science:

• Acoustics and Materials Science
• Computational Physics
• Machine Learning Applications

Background:

• Micro-perforated panels (MPPs) are effective sound absorbers, with performance governed by geometric parameters like porosity and thickness.
• Predicting MPP sound absorption typically involves complex simulations or empirical models.
• Viscous-thermal losses in submillimeter apertures are key to MPP sound absorption.

Purpose of the Study:

• To develop a physics-informed machine learning workflow for predicting the frequency-averaged sound absorption coefficient (α¯) of MPPs.
• To combine acoustically motivated feature engineering with probabilistic prediction and calibrated uncertainty quantification.
• To benchmark various regression pipelines for accuracy and reliability in predicting MPP sound absorption.

Main Methods:

• A quality-controlled dataset of 1000 MPP geometries was created, including parameters like hole radius, porosity, thickness, and perforation shape.
• 28 regression pipelines were benchmarked using leakage-safe fivefold cross-validation.
• A smooth Maa-inspired prior was combined with residual learning, and Gaussian-process (GP) variants were explored, including physics-informed (PI) approaches.

Main Results:

• Physics-informed Gaussian-process (PI-GP) models achieved a root mean square error of 0.138 and R² of 0.763.
• PI-GP posteriors demonstrated near-nominal 95% coverage (0.945), while split-conformal intervals were more conservative (0.970).
• Independent validation using impedance-tube spectra of additively manufactured MPPs showed measured values within the predicted uncertainty intervals.

Conclusions:

• The developed physics-informed machine learning workflow effectively predicts sound absorption in micro-perforated panels.
• The probabilistic predictions offer reliable uncertainty quantification, crucial for practical MPP design.
• While the model shows strong predictive power, discrepancies highlight the need to consider unmodeled effects like cavity, fabrication, and mounting in real-world applications.