26 research outputs found
Designing Multilayered Photonic Shells through Liquid-Liquid Phase Separation in Emulsion Drops
No full text유동-구조 연계를 포함한 효율적 해석을 위한 깊은 계층적 변이 오토인코더의 개발
No full text학위논문(박사) -- 서울대학교대학원 : 공과대학 항공우주공학과, 2024. 2. 신상준.Data-driven model order reduction (MOR) is one of the favorable approaches to accelerate nonlinear dynamic analysis. In this thesis, the development of three data-driven parametric MOR frameworks are presented. The developed frameworks, proper orthogonal decomposition (POD)-Wasserstein generative adversarial network-gradient penalty (WGAN-GP), POD-modified Nouveau variational autoencoder (mNVAE), and least-squares hierarchical variational autoencoder (LSH-VAE) are designed to enhance the accuracy of conventional data-driven parametric MOR frameworks. The first two frameworks utilize POD to reduce the dimensionality of full order model (FOM). POD performs linear MOR, reducing the full degrees of freedom (DOF) to the number of POD modes. Upon the reduced dimension, WGAN-GP and mNVAE are employed to interpolate the POD coefficients. The interpolated POD coefficients are multiplied to the POD modes to generate the interpolated dynamic system. By adopting the latest neural networks, the POD-WGAN-GP and POD-mNVAE gains significantly enhanced accuracy compared to the conventional methods. The WGAN-GP used for the interpolation enhances the accuracy owing to its characteristics of generating ``sharp'' data. mNVAE on the other hand, comprises of deep hierarchical structure, enforcing it to be highly expressive.
While the proposed POD-based methods may be used successfully in many occasions, POD is quite limited as a MOR technique. In cases which show slowly decaying Kolmogorov n-width, POD-based methods require excessive number of POD modes and may exhibit large discrepancy. The LSH-VAE is designed to mitigate such problem by performing nonlinear MOR via its encoder. In the encoder-reduced dimension, spherically linear interpolation is performed for the parametric interpolation. Then, the dimensionality is restored to its original state by the decoder. LSH-VAE exploits three major enhancements to the existing methods: a deep hierarchical structure, a hybrid weighted, probabilistic loss function, and a carefully constructed block constitution. These enhancements enable impressive expressiveness and stability of the network.
The accuracy and efficiency of the proposed methods are demonstrated on various objects. The frameworks are evaluated on five examples, including highly nonlinear, multi-physics, and large number of DOF systems. By adopting the coefficient of determination, the accuracy of the proposed frameworks are evaluated. The proposed frameworks are also compared against the existing pMOR methods such as convolutional autoencoder, Gaussian process regression, and beta-variational autoencoder. As a result, the proposed methods show significantly enhanced accuracy while still exhibiting a large speed-up factor.데이터 기반 차수 축소 기법은 비선형 동적 시스템을 가속화하는 방법 중 하나이다. 본 논문에서는 세 가지의 데이터 기반 차수 축소 기법을 제시하였다. 이 세 가지의 기법은 각각 적합직교분해-Wasserstein 생성형 적대적 네트워크-구배 페널티 (PODWGAN-GP), 적합직교분해-변형 Nouveau 변분 오토 인코더 (POD-mNVAE), 최소 자승 계층적 변분 오토 인코더 (LSH-VAE)이며, 이들은 기존의 비 침습적 파라메트릭 차수 축소 기법의 정확도를 개선하기 위해 고안되었다. 이 중, 첫 두 기법은 적합직교분해를 활용하여 완전 차수 모델의 차원을 축소시킨다. 적합직교분해는 선형 차수축소를
수행하며, 완전 차수 모델 자유도의 개수를 적합직교분해 모드의 개수로 축소시킨다. 축소된 차원에서는 인공신경망인 WGAN-GP와 mNVAE를 활용하여 적합직교분해 계수의 파라메트릭 보간을 수행한다. 보간된 적합직교분해 계수는 적합직교분해 모드와 곱해져 보간된 동적 시스템을 생성한다. 이와 같이 최신의 인공신경망을 도입함으로써, POD-WGAN-GP와 POD-mNVAE는 기존의 기법들 대비 향상된 정확도를 획득한다. 그중, WGAN-GP는 선명한 데이터를 생성한다는 특징을 활용하여 보간의 정확도를 향상시키며 mNVAE는 깊은 계층적 구조로 이루어져 있어, 높은 표현력을 지닌다.
앞서 언급된 적합직교분해 기반의 기법들이 많은 경우에서 효과적으로 사용될 수 있으나 그 효용성은 적합직교분해에 인해 제한된다. 만약 동적 시스템이 느린 Kolmogorovn-width의 붕괴를 가지면 적합직교분해는 과도한 수의 선형모드를 필요로 하며, 큰 오차를 초래할 수 있다. LSH-VAE는 이러한 문제를 완화하기 위해 설계된 기법이며, 비선형 차수 축소 및 보간을 수행한다. LSH-VAE는 축소된 차원에서 구형-선형 보간을 수행하며 그 결과를 본래의 차원으로 복원한다. LSH-VAE는 기존의 신경망 대비 깊은 계층적 구조, 가중, 혼합된 확률론적 손실 함수, 섬세히 설계된 블록 구조라는 세 가지의 주요한 향상점을 가지고 있다. 이런 향상점은 기존 기법 대비 LSH-VAE의 주요한 표현력과 안정성을 가능케 한다.
본 논문에서는 개발된 기법들의 정확도와 효율성은 다양한 대상에 대해 실증하였다. 이 기법들은 강한 비선형성, 다중 물리, 많은 자유도를 가진 시스템을 포함한 다섯 가지 예시를 통해 평가되었다. 또한, 결정 계수를 도입하여 이 기법들을 정량적으로 평가하였으며 합성곱 오토 인코더, 가우시안 프로세스 회귀 법, β-변분 오토 인코더와 같은 기존의 여러 파라메트릭 차수 축소 기법들과 비교하였다. 그 결과, 본 논문에서 제시한 기법들은 기존의 기법들 대비 의미 있는 정확도 향상을 보였으며, 그 동시에 큰 가속인자를 나타내었다Abstract i
Contents iii
List of Figures v
List of Tables viii
1 Introduction 1
1.1 Background and Motivation 1
1.2 Literature Review 6
1.2.1 Classical model order reduction and parametric interpolation 6
1.2.2 Modern model order reduction and parametric interpolation 8
1.3 Objectives 16
1.4 Outline of Dissertation 18
2 Theoretical Background 19
2.1 Governing Equations 19
2.1.1 Structural Dynamics 19
2.1.2 Fluid Dynamics 20
2.1.3 Fluid-Structure Interaction (FSI) 21
2.2 Proper Orthogonal Decomposition (POD) 22
2.3 Artificial Neural Networks 26
2.3.1 Autoencoder 26
2.3.2 Generative Adversarial Network (GAN) 28
2.3.3 Variational Autoencoder (VAE) 36
3 Framework 49
3.1 POD-NN frameworks 49
3.1.1 POD as a MOR method 52
3.1.2 Neural networks as an interpolation method 55
3.2 LSH-VAE 66
4 Numerical Results 73
4.1 2D Stationary Cylinder 74
4.2 Plunging Airfoil 88
4.3 Limit Cycle Oscillation 99
4.4 Turek-Hron FSI 2 case 114
4.5 3D Stationary Cylinder 125
5 Conclusions 134
5.1 Recommendations for Future Works 137
초 록 157박
Asymmetric Pairing of Cholesteric Liquid Crystal Droplets for Programmable Photonic Cross‐Communication
No full textThe photonic cross-communication between photonic droplets has provided complex color patterns through multiple reflections, potentially serving as novel optical codes. However, the cross-communication is mostly restricted to symmetric pairs of identical droplets. Here, a design rule is reported for the asymmetric pairing of two distinct droplets to provide bright color patterns through strong cross-communication and enrich a variety of optical codes. Cholesteric liquid crystal (CLC) droplets with different stopband positions and sizes are paired. The brightness of corresponding color patterns is maximized when the pairs are selected to effectively guide light along the double reflection path by stopbands of two droplets. The experimental results are in good agreement with a geometric model where the blueshift of stopbands is better described by the angles of refraction rather than reflection. The model predicts the effectiveness of pairing quantitatively, which serves as a design rule for programming the asymmetric photonic cross-communication. Moreover, three distinct droplets can be paired in triangular arrays, where all three cross-communication paths yield bright color patterns when three droplets are selected to simultaneously satisfy the rule. It is believed that asymmetric pairing of distinct CLC droplets opens new opportunities for programmable optical encoding in security and anti-counterfeiting applications.
Deep‐Learning‐Based Facial Retargeting Using Local Patches
No full textIn the era of digital animation, the quest to produce lifelike facial animations for virtual characters has led to the development of various retargeting methods. While the retargeting facial motion between models of similar shapes has been very successful, challenges arise when the retargeting is performed on stylized or exaggerated 3D characters that deviate significantly from human facial structures. In this scenario, it is important to consider the target character's facial structure and possible range of motion to preserve the semantics assumed by the original facial motions after the retargeting. To achieve this, we propose a local patch‐based retargeting method that transfers facial animations captured in a source performance video to a target stylized 3D character. Our method consists of three modules. The Automatic Patch Extraction Module extracts local patches from the source video frame. These patches are processed through the Reenactment Module to generate correspondingly re‐enacted target local patches. The Weight Estimation Module calculates the animation parameters for the target character at every frame for the creation of a complete facial animation sequence. Extensive experiments demonstrate that our method can successfully transfer the semantic meaning of source facial expressions to stylized characters with considerable variations in facial feature proportion.</jats:p>
ASCAT2SMAP: Image-to-Image Translation to Obtain L-Band-Like Soil Moisture From C-Band Satellite Data
No full textSoil moisture (SM) is a critical parameter in understanding the Earth's hydrological cycle and managing water resources. Remote sensing instruments, such as Advanced SCATterometer (ASCAT), can provide valuable long-term SM. However, compatibility issues may arise when integrating ASCAT SM retrieval with another retrieval, such as soil moisture active passive (SMAP), a high-quality microwave radiometer-based SM retrieval. In this study, we propose a novel image-to-image translation approach based on the U-Net architecture to convert ASCAT SM data into the format of SMAP (ASCAT2SMAP). The resulting SM from the ASCAT2SMAP was evaluated using temporally separated SMAP data and independent in-situ SM measurement from the International Soil Moisture Network (ISMN). In the separately divided test periods, ASCAT2SMAP showed good agreement with SMAP with R of 928, ubRMSD of 0.043 , and bias of 0.002 . When evaluating ASCAT2SMAP with ISMN data, it showed a better agreement than ASCAT and more similar metrics with SMAP. Moreover, we found that the ASCAT2SMAP is more robust to a problem of subsurface scattering than the original ASCAT SM. When simulating V-polarized brightness temperature from ASCAT2SMAP SM, it showed good agreement with ubRMSD of 5.602 K and bias of −0.135 K. Our results are expected to provide a valuable perspective preceding to creation of harmonized SM datasets from different sensors, contributing to improved data integration and analysis in the field of geoscience and remote sensing
Tumble Flow Measurements Using Three Different Methods and its Effects on Fuel Economy and Emissions
No full textTumble Flow Measurements Using Three Different Methods and Its Effects on Fuel Economy and Emissions
No full textV advancement eversion flap for fingertip injury: Preventing ischemia and hook-nail deformity
No full textIntroduction: Traumatic fingertip amputation is the most common type of upper extremity injuries. The V-Y advancement flap is a reliable method for reconstructing fingertip defects, but it is associated with complications such as hook-nail deformity and suture site ischemia. Here, we describe our modifications to V-Y advancement flap technique, termed as “V advancement eversion flap” and review the outcomes of this procedure in 21 patients with fingertip amputation. Methods: This was a retrospective review of 21 consecutive patients with fingertip injury who were treated surgically using the V advancement eversion flap technique at a single trauma center between 2006 and 2019. We analyzed the age, injury location and mechanism, Allen classification, injury geometry, and objective and subjective clinical outcomes. Results: Twenty-three fingertip amputations with defect sizes greater than 1.0 cm2 from the tip to lunula were included in this study. The mean age of the patients was 43.6 years (range, 24–65 years). The average follow-up period was 20 months (range, 12–37 months). The average wound healing time (apparent epithelization) was 29.4 days (range, 14-41 days). At the final follow-up, all flaps had healed uneventfully without noticeable hook-nail deformity. In the static two-point discrimination test, the mean value was 4.61 mm in the injured finger. Patient ratings of the outcomes were “excellent” in 18 and “good” in 5 cases. Conclusion: The V advancement eversion flap technique, when properly designed and executed in fingertip amputation cases, can minimize morbidity and result in successful wound healing without flap necrosis and hook-nail deformity
Data-driven Nonlinear Parametric Model Order Reduction Framework using Deep Hierarchical Variational Autoencoder
No full textA data-driven parametric model order reduction (MOR) method using a deep
artificial neural network is proposed. The present network, which is the
least-squares hierarchical variational autoencoder (LSH-VAE), is capable of
performing nonlinear MOR for the parametric interpolation of a nonlinear
dynamic system with a significant number of degrees of freedom. LSH-VAE
exploits two major changes to the existing networks: a hierarchical deep
structure and a hybrid weighted, probabilistic loss function. The enhancements
result in a significantly improved accuracy and stability compared against the
conventional nonlinear MOR methods, autoencoder, and variational autoencoder.
Upon LSH-VAE, a parametric MOR framework is presented based on the spherically
linear interpolation of the latent manifold. The present framework is validated
and evaluated on three nonlinear and multiphysics dynamic systems. First, the
present framework is evaluated on the fluid-structure interaction benchmark
problem to assess its efficiency and accuracy. Then, a highly nonlinear
aeroelastic phenomenon, limit cycle oscillation, is analyzed. Finally, the
present framework is applied to a three-dimensional fluid flow to demonstrate
its capability of efficiently analyzing a significantly large number of degrees
of freedom. The performance of LSH-VAE is emphasized by comparing its results
against that of the widely used nonlinear MOR methods, convolutional
autoencoder, and -VAE. The present framework exhibits a significantly
enhanced accuracy to the conventional methods while still exhibiting a large
speed-up factor
Experimental study of micro patterns on conductive ink transfer in a plate-to-roll system
No full text
