Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor.advisor | 노준혁 | - |
| dc.contributor.author | 황서영 | - |
| dc.creator | 황서영 | - |
| dc.date.accessioned | 2026-02-23T16:32:31Z | - |
| dc.date.available | 2026-02-23T16:32:31Z | - |
| dc.date.issued | 2026 | - |
| dc.identifier.other | OAK-000000259512 | - |
| dc.identifier.uri | https://dcollection.ewha.ac.kr/common/orgView/000000259512 | en_US |
| dc.identifier.uri | https://dspace.ewha.ac.kr/handle/2015.oak/274475 | - |
| dc.description.abstract | Predicting subject-specific foot deformation under ground contact is essential for ergonomic footwear design and clinical biomechanics, yet traditional Finite Element Method (FEM) simulations remain computationally expensive for real-time analysis or iterative design. While FEM offers high physical fidelity, its cost limits scalability. Recent studies have explored Graph Neural Networks (GNNs), particularly MeshGraphNet (MGN), as surrogate models for mesh-based simulations; however, existing frameworks typically underexploit inter-subject biological variability and are prone to temporal error accumulation during multi-step rollout. This thesis introduces Bio-Adaptive MeshGraphNet, a bio-informed mesh simulation framework built upon MGN for foot deformation prediction under ground contact. Biological metadata is incorporated through node-wise conditioning (Bio-Node), improving generalization across subjects with varying anatomical characteristics. To enhance temporal robustness, we employ multi-step rollout training, and further enforce physical plausibility using continuum-mechanics-inspired regularizers whose weights are adaptively inferred from a biological embedding via a Bio-Material Parameter Network. In addition, an adaptive node masking strategy is proposed to mitigate artifacts near boundary and contact regions. Evaluations on a high-fidelity Abaqus-based dataset indicate that Bio-Adaptive MeshGraphNet improves prediction accuracy and deformation stability compared to representative GNN and Transformer baselines, particularly in high-contact regions. These findings further provide empirical insight into how biological priors interact with physical regularization, highlighting their importance in contact-driven deformation modeling.;지면 접촉 상황에서 피험자별 발 변형을 정확히 예측하는 기술은 인체공학적 신발 설계와 임상 생체역학 분야에서 필수적이다. 그러나 기존 Finite Element Method(FEM) 기반 시뮬레이션은 높은 물리적 분석이나 반복적 설계 과정에 적용하기 어렵다. 이를 해결하기 위해 최근 Graph Neural Network(GNN), 특히 MeshGraphNet(MGN)을 활용한 메쉬 기반 시뮬레이션 대체 모델이 연구되고 있으나, 기존 방법은 피험자 간 생체 정보 변이를 충분히 반영하지 못하고, 다중 스텝 rollout 과정에서 예측 오차가 누적되는 한계가 있다. 본 논문은 이러한 문제를 해결하기 위해 Bio-Adaptive MeshGraphNet을 제안한다. 제안 모델은 생체 메타데이터를 노드 수준에서 조건부 주입하는 Bio-Node 방식을 통해 피험자별 해부학적 차이를 반영하고 일반화 성능을 향상시킨다. 또한 multi-step rollout training을 적용하여 장기 시뮬레이션의 안정성을 높이고, 연속체 역학 기반 정규화 항을 도입하여 물리적으로 일관된 변형을 유도한다. 이때 정규화 가중치는 Bio-Material Parameter Network를 통해 bio-vector로부터 적응적으로 산출되어 개인별 재료 특성을 반영한다. 아울러 경계 및 접촉 영역에서 발생할 수 있는 artifact를 완화하기 위해 adaptive node masking 전략을 도입하였다. Abaqus 기반 고정밀 데이터셋을 활용한 평가 결과, Bio-Adaptive MeshGraphNet은 대표적인 GNN 및 Transformer 기반 모델 대비 예측 정확도와 변형 안정성에서 개선된 성능을 보였으며, 특히 고접촉 영역에서 피크 변형을 더 안정적으로 포착하였다. 또한 생체 prior와 물리 정규화 항의 상호작용을 실험적으로 분석하여, contact 기반 변형 모델링에서 두 요소의 균형적 결합이 중요함을 확인하였다. | - |
| dc.description.tableofcontents | I. Introduction 1 II. Related Works 3 1. Learning Mesh-Based Simulation 3 1.1. General Models 3 1.2. Mesh-based Deformation Prediction 3 2. Graph Neural Networks 4 2.1. General Graph Neural Networks 4 2.2. Graph Transformers in Physics 5 3. Physics-Informed Deep Learning 5 4. Bio-conditioned and Physics-Constrained Surrogates 6 III. Our Dataset 7 1. Dataset Acquisition 7 IV. Methodology 9 1. Problem Formulation and Graph Representation 9 2. Bio-Informed Architecture 10 2.1. Biological Context Injection (Bio-Node) 10 2.2. Deep Unnormalized Topology 11 3. Multi-Step Rollout Training 12 4. Adaptive Physical Regularization and Loss 13 4.1. Geometric Smoothing Regularizers 14 4.2. As-Rigid-As-Possible (ARAP) Constraint 14 4.3. Neo-Hookean Hyperelasticity 15 4.4. Topological Consistency Loss 15 5. Total Training Objective 16 6. Architectural Comparison to Transformer Baselines 17 V. Experiments 19 1. Experimental Setup 19 1.1. Implementation Details 19 2. Dataset Setting 20 3. Evaluation Metrics and Baseline 20 3.1. Zero-Displacement Baseline 21 3.2. Evaluation Metrics 21 4. Quantitative Analysis 23 5. Qualitative Analysis 25 6. Deformation Overlay Analysis 27 VI. Discussion 28 1. Performance Superiority 28 1.1. Axis-wise performance 28 1.2. Interpretation of RMSE-MAE Discrepancy 29 2. Ablation Analysis 30 3. Limitations and Future Directions 31 VII. Conclusion 32 Bibliography 33 Appendix 41 | - |
| dc.format | application/pdf | - |
| dc.format.extent | 6570905 bytes | - |
| dc.language | eng | - |
| dc.publisher | 이화여자대학교 대학원 | - |
| dc.subject | Graph Neural Networks, MeshGraphNet, Bio-Adaptive Simulation, Foot Deformation, Biomechanics, Physics-Informed Learning, Rollout Training | - |
| dc.subject.ddc | 000 | - |
| dc.title | Bio-Adaptive MeshGraphNet for High-Fidelity Foot Deformation Prediction under Ground Contact | - |
| dc.type | Master's Thesis | - |
| dc.creator.othername | Hwang, Seoyeong | - |
| dc.format.page | vi, 41 p. | - |
| dc.identifier.thesisdegree | Master | - |
| dc.identifier.major | 대학원 인공지능·소프트웨어학부 | - |
| dc.date.awarded | 2026. 2 | - |