Graph sample and aggregate-based deep operator network for accelerating one-dimensional human breathing simulations.

Area of Science:

• Computational fluid dynamics
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• Neural network (NN)-based surrogate models offer potential for accelerating human lung airflow simulations.
• The lung airway's hierarchical structure and complex geometry present challenges for NN model design due to wide flow rate variations.

Purpose of the Study:

• To introduce a novel GraphDeepONet model for accelerating 1D human airway simulations.
• To develop a volume-based normalization technique to handle multiscale flow rate variations.
• To compare the performance of three different GraphDeepONet model architectures.

Main Methods:

• Introduced GraphDeepONet, combining Deep Operator Network (DeepONet) and Graph Sample and Aggregate (GraphSAGE).
• Implemented a volume-based normalization technique for training stability.
• Evaluated three models: dual-model prediction, sequential prediction, and hierarchical acinar region prediction.

Main Results:

• Model 3, using hierarchical acinar region prediction, showed the highest accuracy for healthy subjects (e.g., 7.0% L2 relative error in flow rate).
• GraphDeepONet models achieved significant speedups, reducing computation time from ~12 minutes to <1 second.
• Model 3 demonstrated median errors of 7.0% for flow rate, 5.9% for static pressure, and 6.8% for pleural pressure.

Conclusions:

• The GraphDeepONet surrogate modeling strategy enables efficient simulation of 1D human airway airflow.
• This approach has the potential to facilitate real-time respiratory analysis.
• Personalized treatment strategies can be enhanced through rapid and accurate airflow simulations.