An Operator Learning Approach to Efficient 1D Airflow Simulation in the Lungs : 폐의 효율적인 1D 공기 흐름 시뮬레이션을 위한 운영자 학습 접근 방식

다국어 초록 (Multilingual Abstract)

The application of computational fluid dynamics (CFD) in simulating pulmonary breathing plays a vital role in advancing the understanding of lung physiology and pathology. Beyond providing mechanistic insights, CFD-based models enable the exploration of multiple physiological and pathological scenarios at relatively low cost, thereby reducing the dependence on physical prototypes when boundary conditions change. Despite these advantages, the complex branching structure of the airways introduces challenges in mesh generation, often requiring expert intervention to ensure a high-quality mesh suitable for CFD simulations. Furthermore, dense meshes, while necessary to capture the intricate airway geometry, significantly increase computational demands, thus limiting the feasibility of real- time, patient-specific simulations .
To overcome these challenges, this thesis proposes a Deep Operator Network based on Graph Neural Network, namely GraphDeepONet, to develop a surrogate model for the 1D airway breathing simulations. The surrogate model is trained on a benchmark dataset comprising flow rate, pressure, and pleural pressure. Given the hierarchical airway geometry that induces multiscale flow variations, we propose a novel volume-based normalization technique to mitigate these effects during training.
Additionally, three distinct learning strategies are introduced based on GraphDeepONet. Model 1 employs a dual-network GraphDeepONet to predict flow rate and pressure at the same time with physical constraint. Model 2 first predicts flow rate and computes pressure based on the predicted flow rate in the training phase. Similar to Model 2 in computing pressure, but in flow rate prediction, Model 3 first predicts the flow rate at the acinar region, then summing up to obtain the other flow rate. Among them, Model 3 achieved a prediction L2 relative error of 10.44% (5.08%) in flow rate, 9.99% (5.52%) in pressure, and 9.29% (4.10%) in pleural pressure.
Overall, the proposed method significantly speeds up computational time at least 700x compared to conventional solvers. This work presents an efficient surrogate model for 1D airway breathing simulation, offering a potential application for real-time respiratory analysis and the development of personalized treatment strategies.

목차 (Table of Contents)

• CHAPTER 1: INTRODUCTION 1
• 1.1 Backgrounds 1
• 1.1.1 Respiratory system 1
• 1.1.2 Physiological phenomena of the Lung 3
• 1.1.3 Common methods for detecting lung diseases 4
• 1.1.4 Computational fluid dynamics 6
• 1.1.5 Deep Learning 9
• 1.2 Thesis objectives 14
• CHAPTER 2: Deep Learning-based Surrogate Modeling for One-Dimensional Human Airway Flow Simulation 16
• 2.1 Reduced-order 1D airway flows 16
• 2.1.1 Governing equation 16
• 2.1.2 Data acquisition 18
• 2.2 Methodology 19
• 2.2.1 GraphDeepONet 19
• 2.2.2 Data normalization 23
• 2.2.3 Loss function 25
• 2.3 Results 26
• 2.3.1 Physiological distribution 26
• 2.3.2 Hysteresis curve 30
• 2.3.3 Quantitative evaluation 31
• 2.3.4 Evaluation of Mass Conservation 33
• 2.3.5 Computational cost 33
• 2.4 Discussion 34
• 2.4.1 Different backbones for the branch network 34
• 2.4.2 Novel volume-based normalization on training 34
• 2.4.3 Strengths and Limitations 36
• CHAPTER 3: SUMMARY AND FUTURE WORKS 38
• REFERENCES 39