Physics-informed neural network for predicting in vacuo vocal fold eigenmodes: A proof of concept study

Area of Science:

• Computational physics
• Machine learning in biomechanics
• Vocal fold dynamics

Background:

• Accurate computation of vocal fold eigenmodes is crucial for understanding voice production.
• Traditional methods like finite-element analysis can be computationally intensive.
• Real-time prediction methods are needed for dynamic simulations and clinical applications.

Purpose of the Study:

• To develop and evaluate a machine learning approach for real-time computation of in vacuo vocal fold eigenmodes.
• To integrate governing equations of vocal fold dynamics into a physics-informed neural network.
• To assess the accuracy and efficiency of the proposed method compared to finite-element results.

Main Methods:

• Training a physics-informed neural network (PINN) to predict vocal fold eigenmodes and eigenfrequencies.
• Integrating the governing equations of vocal fold dynamics into the neural network architecture.
• Comparing PINN predictions with results from finite-element method (FEM) simulations.

Main Results:

• The PINN accurately predicted physically consistent modal shapes and eigenfrequency estimates.
• Strong agreement was observed between PINN results and FEM for lower-order modes.
• Mean relative errors below 6% in eigenfrequency prediction and cosine correlation values near 1 for eigenmodes were achieved.
• Prediction accuracy decreased for higher-order modes.

Conclusions:

• Physics-informed neural networks offer an accurate and efficient method for real-time computation of vocal fold eigenmodes.
• This approach holds potential for advancing voice production research and clinical diagnostics.
• Further refinement may improve accuracy for higher-order modes.