A Deep Neural Network Based Glottal Flow Model for Predicting Fluid-Structure Interactions during Voice Production.

Area of Science:

• Computational fluid dynamics
• Voice production biomechanics
• Machine learning in biomechanics

Background:

• Accurate glottal flow modeling is crucial for understanding voice production.
• Traditional methods like Navier-Stokes simulations are computationally expensive.
• Reduced-order models offer a balance between accuracy and efficiency.

Purpose of the Study:

• To develop a fast and accurate machine learning-based reduced-order model for glottal flow prediction.
• To integrate a deep neural network (DNN) with the Bernoulli equation for improved viscous loss prediction.
• To enable efficient simulation of fluid-structure interactions (FSI) in voice production.

Main Methods:

• A deep neural network (DNN) was trained on 3D Navier-Stokes (N-S) simulations of glottal flows for various glottal shapes.
• The DNN predicts viscous loss terms, integrated into a Bernoulli equation-based model.
• The model was coupled with a finite-element method solver for fluid-structure interaction (FSI) simulations.

Main Results:

• The DNN-Bernoulli model accurately predicts flow resistance, flow rate, and pressure distribution for diverse glottal shapes.
• The model demonstrates good prediction performance in both static shape and FSI simulations.
• Computational efficiency was significantly improved compared to traditional N-S models.

Conclusions:

• The developed machine learning model provides accurate and efficient predictions of glottal flow.
• The model shows significant promise for future clinical applications in voice analysis and treatment.
• This approach advances the computational modeling of voice production biomechanics.