본문

서브메뉴

Methods for Learning Mechanics: Inverse Problems, Constitutive Modeling, and Design
Methods for Learning Mechanics: Inverse Problems, Constitutive Modeling, and Design
Methods for Learning Mechanics: Inverse Problems, Constitutive Modeling, and Design

상세정보

자료유형  
 학위논문 서양
최종처리일시  
20260202104751
ISBN  
9798290655758
DDC  
300
저자명  
Rajan, Aakila.
서명/저자  
Methods for Learning Mechanics: Inverse Problems, Constitutive Modeling, and Design
발행사항  
[Sl] : California Institute of Technology, 2025
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2025
형태사항  
206 p
주기사항  
Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
주기사항  
Advisor: Bhattacharya, Kaushik.
학위논문주기  
Thesis (Ph.D.)--California Institute of Technology, 2025.
초록/해제  
요약Predicting the behavior of physical systems under complex loading conditions remains a central challenge in engineering and the applied sciences. In solid mechanics, accurate prediction requires not only knowledge of the governing physical laws but also appropriate constitutive models that describe material behavior. These constitutive models are parameterized functions typically obtained through material characterization techniques. However, traditional techniques have limited applicability due to their reliance on challenging experiments designed to produce homogeneous fields, especially for complex materials.First, we propose a method to accurately and efficiently identify the constitutive behavior of complex materials from full-field observations. We formulate the problem of inferring constitutive relations as an indirect inverse problem constrained by the balance laws. Specifically, we seek a constitutive relation that minimizes the difference between experimental observations and model predictions while strictly enforcing physical laws. The forward problem is posed as a boundary value problem corresponding to the experiment, and sensitivities are computed using the adjoint method. The resulting framework is robust and applicable to constitutive models of arbitrary complexity. We focus on elasto-viscoplasticity and demonstrate the method using synthetic data on two problems, one quasistatic and the other dynamic.We then extend this methodology to infer constitutive parameters from experimental dynamic contact data, where challenges such as noise, incomplete information, and model-form uncertainties complicate inversion. Using force-depth measurements from split Hopkinson pressure bar experiments on steel and aluminum samples, and incorporating regularization techniques to mitigate illposedness, the framework robustly recovers material properties. Results show the successful extraction of meaningful parameters, highlighting the method's effectiveness over traditional direct inversion approaches, particularly for nonlinear models.A key limitation of the proposed method is its reliance on a pre-specified form of the constitutive model. The choice of such forms becomes harder with the evergrowing catalog of materials. To address this, we develop a neural network-based framework where the constitutive relation emerges directly from data. Using a recurrent neural operator to model history-dependent behavior with internal variables, we demonstrate on synthetic high strain rate compression experiments that this approach significantly outperforms traditional models in capturing path-dependent material responses. Therefore, this proves that neural networks can be powerful tools for constitutive modeling.Building on the ability of neural networks to approximate complex functions, we further explore their use in approximating solutions to partial differential equations (PDEs). In following study, we investigate neural operators for multiscale modeling of elliptic PDEs, focusing on learning effective macroscopic behavior from complex microstructures without repeatedly solving finescale problems. Targeting the classical homogenization setting of linear elliptic PDEs with discontinuous coefficients, we analyze the challenges posed by sharp interfaces and degraded solution regularity. We provide theoretical guarantees on approximation capabilities and demonstrate the method across various representative microstructures, showing that neural operator-based homogenization offers a scalable and non-intrusive approach to multiscale modeling.Finally, in the last study, we extend these data-driven tools to topology optimization, where the need for numerous PDE solves presents a major computational bottleneck. By integrating a reduced-order PCA-based neural network into the optimization loop, we represent complex structures in a lowdimensional latent space and achieve efficient, high-quality updates. This approach significantly accelerates optimization while preserving design flexibility, highlighting the potential of machine learning to enhance and transform classical design methodologies.Together, these contributions present a unified strategy that integrates physicsbased modeling, machine learning, and optimization to advance material characterization, multiscale modeling, and design in solid mechanics.
일반주제명  
Load
일반주제명  
Sensitivity analysis
일반주제명  
Parameter identification
일반주제명  
Neural networks
일반주제명  
Homogenization
일반주제명  
Microstructure
일반주제명  
Mechanics
일반주제명  
Boundary conditions
일반주제명  
Geometry
기타저자  
California Institute of Technology Engineering and Applied Science
기본자료저록  
Dissertations Abstracts International. 87-01B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008260126s2025        us                              c    eng  d
■001000017358781
■00520260202104751
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798290655758
■035    ▼a(MiAaPQ)AAI32151340
■035    ▼a(MiAaPQ)Caltech17272
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a300
■1001  ▼aRajan,  Aakila.
■24510▼aMethods  for  Learning  Mechanics:  Inverse  Problems,  Constitutive  Modeling,  and  Design
■260    ▼a[Sl]▼bCalifornia  Institute  of  Technology▼c2025
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2025
■300    ▼a206  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  87-01,  Section:  B.
■500    ▼aAdvisor:  Bhattacharya,  Kaushik.
■5021  ▼aThesis  (Ph.D.)--California  Institute  of  Technology,  2025.
■520    ▼aPredicting  the  behavior  of  physical  systems  under  complex  loading  conditions  remains  a  central  challenge  in  engineering  and  the  applied  sciences.  In  solid  mechanics,  accurate  prediction  requires  not  only  knowledge  of  the  governing  physical  laws  but  also  appropriate  constitutive  models  that  describe  material  behavior.  These  constitutive  models  are  parameterized  functions  typically  obtained  through  material  characterization  techniques.  However,  traditional  techniques  have  limited  applicability  due  to  their  reliance  on  challenging  experiments  designed  to  produce  homogeneous  fields,  especially  for  complex  materials.First,  we  propose  a  method  to  accurately  and  efficiently  identify  the  constitutive  behavior  of  complex  materials  from  full-field  observations.  We  formulate  the  problem  of  inferring  constitutive  relations  as  an  indirect  inverse  problem  constrained  by  the  balance  laws.  Specifically,  we  seek  a  constitutive  relation  that  minimizes  the  difference  between  experimental  observations  and  model  predictions  while  strictly  enforcing  physical  laws.  The  forward  problem  is  posed  as  a  boundary  value  problem  corresponding  to  the  experiment,  and  sensitivities  are  computed  using  the  adjoint  method.  The  resulting  framework  is  robust  and  applicable  to  constitutive  models  of  arbitrary  complexity.  We  focus  on  elasto-viscoplasticity  and  demonstrate  the  method  using  synthetic  data  on  two  problems,  one  quasistatic  and  the  other  dynamic.We  then  extend  this  methodology  to  infer  constitutive  parameters  from  experimental  dynamic  contact  data,  where  challenges  such  as  noise,  incomplete  information,  and  model-form  uncertainties  complicate  inversion.  Using  force-depth  measurements  from  split  Hopkinson  pressure  bar  experiments  on  steel  and  aluminum  samples,  and  incorporating  regularization  techniques  to  mitigate  illposedness,  the  framework  robustly  recovers  material  properties.  Results  show  the  successful  extraction  of  meaningful  parameters,  highlighting  the  method's  effectiveness  over  traditional  direct  inversion  approaches,  particularly  for  nonlinear  models.A  key  limitation  of  the  proposed  method  is  its  reliance  on  a  pre-specified  form  of  the  constitutive  model.  The  choice  of  such  forms  becomes  harder  with  the  evergrowing  catalog  of  materials.  To  address  this,  we  develop  a  neural  network-based  framework  where  the  constitutive  relation  emerges  directly  from  data.  Using  a  recurrent  neural  operator  to  model  history-dependent  behavior  with  internal  variables,  we  demonstrate  on  synthetic  high  strain  rate  compression  experiments  that  this  approach  significantly  outperforms  traditional  models  in  capturing  path-dependent  material  responses.  Therefore,  this  proves  that  neural  networks  can  be  powerful  tools  for  constitutive  modeling.Building  on  the  ability  of  neural  networks  to  approximate  complex  functions,  we  further  explore  their  use  in  approximating  solutions  to  partial  differential  equations  (PDEs).  In  following  study,  we  investigate  neural  operators  for  multiscale  modeling  of  elliptic  PDEs,  focusing  on  learning  effective  macroscopic  behavior  from  complex  microstructures  without  repeatedly  solving  finescale  problems.  Targeting  the  classical  homogenization  setting  of  linear  elliptic  PDEs  with  discontinuous  coefficients,  we  analyze  the  challenges  posed  by  sharp  interfaces  and  degraded  solution  regularity.  We  provide  theoretical  guarantees  on  approximation  capabilities  and  demonstrate  the  method  across  various  representative  microstructures,  showing  that  neural  operator-based  homogenization  offers  a  scalable  and  non-intrusive  approach  to  multiscale  modeling.Finally,  in  the  last  study,  we  extend  these  data-driven  tools  to  topology  optimization,  where  the  need  for  numerous  PDE  solves  presents  a  major  computational  bottleneck.  By  integrating  a  reduced-order  PCA-based  neural  network  into  the  optimization  loop,  we  represent  complex  structures  in  a  lowdimensional  latent  space  and  achieve  efficient,  high-quality  updates.  This  approach  significantly  accelerates  optimization  while  preserving  design  flexibility,  highlighting  the  potential  of  machine  learning  to  enhance  and  transform  classical  design  methodologies.Together,  these  contributions  present  a  unified  strategy  that  integrates  physicsbased  modeling,  machine  learning,  and  optimization  to  advance  material  characterization,  multiscale  modeling,  and  design  in  solid  mechanics.
■590    ▼aSchool  code:  0037.
■650  4▼aLoad
■650  4▼aSensitivity  analysis
■650  4▼aParameter  identification
■650  4▼aNeural  networks
■650  4▼aHomogenization
■650  4▼aMicrostructure
■650  4▼aMechanics
■650  4▼aBoundary  conditions
■650  4▼aGeometry
■690    ▼a0346
■71020▼aCalifornia  Institute  of  Technology▼bEngineering  and  Applied  Science.
■7730  ▼tDissertations  Abstracts  International▼g87-01B.
■790    ▼a0037
■791    ▼aPh.D.
■792    ▼a2025
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358781▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

미리보기

내보내기

chatGPT토론

Ai 추천 관련 도서


    신착도서 더보기
    최근 3년간 통계입니다.

    소장정보

    • 예약
    • 소재불명신고
    • 나의폴더
    • 우선정리요청
    • 비도서대출신청
    • 야간 도서대출신청
    소장자료
    등록번호 청구기호 소장처 대출가능여부 대출정보
    TF19405 전자도서 대출가능 마이폴더 부재도서신고 비도서대출신청 야간 도서대출신청

    * 대출중인 자료에 한하여 예약이 가능합니다. 예약을 원하시면 예약버튼을 클릭하십시오.

    해당 도서를 다른 이용자가 함께 대출한 도서

    관련 인기도서

    로그인 후 이용 가능합니다.