Objective: This study aims to develop and validate a fast intraglottal pressure prediction model using Diffeomorphic Mapping Operator Learning (DIMON), a geometry-aware neural operator. The model can instantly predict surface pressure for arbitrary glottal geometries, including those with left–right asymmetry and multiple flow channels. Once developed, DIMON can be coupled with finite element–based vocal fold simulations to provide fast, high-fidelity predictions across normal and diseased phonatory conditions without solving the computationally expensive Navier–Stokes equations.
Background: Precise wall pressure inside the glottis is critical for modeling vocal fold vibrations. Bernoulli-based methods fail when vocal folds contact and form asymmetric glottal shapes and multiple flow channels, while full 3D fluid–structure simulations, though accurate, are computationally prohibitive. Neural predictors are faster but often struggle to generalize across arbitrary geometries. Recently, DIMON has been introduced to address this limitation by mapping diverse geometries to a common reference domain, learning a unified pressure operator, and transferring the solution back to the original shape. This enables fast and accurate pressure prediction across varied glottal configurations.
Methods: We generated a collection of vocal fold geometries by pairing left and right vibration patterns using Titze’s surface-wave model. Each vibration cycle was sampled at 32 equally spaced phases, treating each phase as a distinct static geometry. For each geometry, we performed Navier–Stokes simulations using an immersed boundary method to obtain steady-state wall-pressure fields. These geometries and their corresponding pressure fields comprise the training dataset for DIMON. New vibration patterns are iteratively added until DIMON accurately predicts previously unseen cases.
Results and Conclusions: With a limited training set (5 vibration patterns × 32 phases), DIMON was able to roughly predict the intraglottal pressure fields for previously unseen vocal fold geometries. The model’s accuracy improves as the number and diversity of training cases increase. We plan to expand the dataset with additional CFD cases to further enhance prediction fidelity. Comparison with CFD results demonstrates DIMON’s potential to replace costly per-phase CFD calculations while maintaining accurate wall-pressure predictions across varied glottal geometries