東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

Surrogate Modeling and Inverse Analy...

Yang, Qinghao,

FindBook

Google Book

Amazon

博客來

Surrogate Modeling and Inverse Analysis for Soil Plasticity With Machine Learning /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Surrogate Modeling and Inverse Analysis for Soil Plasticity With Machine Learning // Qinghao Yang.
作者:	Yang, Qinghao,
面頁冊數:	1 electronic resource (153 pages)
附註:	Source: Dissertations Abstracts International, Volume: 85-10, Section: B.
Contained By:	Dissertations Abstracts International85-10B.
標題:	Geotechnology. -
電子資源:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30993550
ISBN:	9798381975284

Surrogate Modeling and Inverse Analysis for Soil Plasticity With Machine Learning /
Yang, Qinghao,

Surrogate Modeling and Inverse Analysis for Soil Plasticity With Machine Learning /Qinghao Yang. - 1 electronic resource (153 pages)

Source: Dissertations Abstracts International, Volume: 85-10, Section: B.

In engineering design or inverse analysis, it is usually required to rerun experiments or simulations to find the optimal parameters. The multiple runs of experiments or simulations are time-consuming for complex engineering problems. An efficient alternative is building surrogate models between quantities of interest and corresponding parameters. Recently, machine learning has been applied widely to construct surrogate models in engineering, including geotechnical engineering. However, little research incorporated the large deformation of soils and the corresponding evolution of soil geometries when building machine-learning-based surrogate models, although large deformation occurs in many geotechnical applications. Thus, the first part of this work is building surrogate models that take account of large deformation. Two problems are studied in the first part.The first problem addressed in this work pertains to 2D granular column collapse. A data-driven surrogate model is introduced and validated for predicting the moving geometries during column collapse. The trained surrogate model can accurately predict the movements of granular columns with initial geometries not present in the trained data. Two frameworks, namely a Lagrangian framework and an Eulerian framework, are proposed to establish rules for data collection. Results indicate that the Eulerian framework can give a better prediction. Additionally, a sensitivity analysis is conducted to explore the impact of time increment and grid spacing.The second problem concerns soil-structure/tool interaction. A neural network-based surrogate model is developed to learn the complete history of force-displacement relations during soil-structure/tool interaction. To handle the evolution of free surfaces caused by large deformation, the geometries of free surfaces serve as both inputs and outputs of the neural network. To mitigate the "curse of dimensionality", the dimensions representing a free surface are reduced before training the neural network. The model is tested on three fundamental problems involving large deformation: wedge indentation, cylinder indentation, and plate ploughing. Comparison with results from existing numerical methods demonstrates that the surrogate model can predict responses for various geometries of tools, soil parameters, and moving paths.The second part of the work focuses on evaluating the effectiveness and efficiency of physics-informed neural networks (PINNs) in inverse analysis for soil plasticity. Compared to inverse analysis with surrogate models built through black-box machine learning, PINNs have the advantage of requiring significantly less data. A PINN is developed specifically for elastoplastic materials. To capture the history-dependent properties of such materials, a recurrent neural network (RNN) is used. Before engaging in inverse analysis, the proposed model is validated in forward analysis through two typical tests for soils, i.e., biaxial compression and flat punch indentation. Results from the model closely align with finite element solutions. In terms of inverse analysis, the model successfully identifies plastic parameters of elastoplastic materials without the need for iterative simulations. To eliminate some efforts in manually adjusting the loss function of the neural network, a hard imposition of measurements is employed, automatically ensuring that the outputs of the neural network adhere to the measured data. To accelerate the inverse analysis process, different strategies of transfer learning are also explored.

English

ISBN: 9798381975284Subjects--Topical Terms:

1018558
Geotechnology.
Subjects--Index Terms:

Inverse analysis

Surrogate Modeling and Inverse Analysis for Soil Plasticity With Machine Learning /
LDR:05091nmm a22004573i 4500 001 2428521
005 20250414021845.5
006 m o d
007 cr|nu||||||||
008 261011s2024 miu||||||m |||||||eng d
020 $a 9798381975284
035 $a (MiAaPQD)AAI30993550
035 $a AAI30993550
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Yang, Qinghao, $e author. $3 3815790
245 1 0 $a Surrogate Modeling and Inverse Analysis for Soil Plasticity With Machine Learning / $c Qinghao Yang.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 1 electronic resource (153 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-10, Section: B.
500 $a Advisors: Hambleton, James Committee members: Buscarnera, Giuseppe; Rotta Loria, Alessandro.
502 $b Ph.D. $c Northwestern University $d 2024.
520 $a In engineering design or inverse analysis, it is usually required to rerun experiments or simulations to find the optimal parameters. The multiple runs of experiments or simulations are time-consuming for complex engineering problems. An efficient alternative is building surrogate models between quantities of interest and corresponding parameters. Recently, machine learning has been applied widely to construct surrogate models in engineering, including geotechnical engineering. However, little research incorporated the large deformation of soils and the corresponding evolution of soil geometries when building machine-learning-based surrogate models, although large deformation occurs in many geotechnical applications. Thus, the first part of this work is building surrogate models that take account of large deformation. Two problems are studied in the first part.The first problem addressed in this work pertains to 2D granular column collapse. A data-driven surrogate model is introduced and validated for predicting the moving geometries during column collapse. The trained surrogate model can accurately predict the movements of granular columns with initial geometries not present in the trained data. Two frameworks, namely a Lagrangian framework and an Eulerian framework, are proposed to establish rules for data collection. Results indicate that the Eulerian framework can give a better prediction. Additionally, a sensitivity analysis is conducted to explore the impact of time increment and grid spacing.The second problem concerns soil-structure/tool interaction. A neural network-based surrogate model is developed to learn the complete history of force-displacement relations during soil-structure/tool interaction. To handle the evolution of free surfaces caused by large deformation, the geometries of free surfaces serve as both inputs and outputs of the neural network. To mitigate the "curse of dimensionality", the dimensions representing a free surface are reduced before training the neural network. The model is tested on three fundamental problems involving large deformation: wedge indentation, cylinder indentation, and plate ploughing. Comparison with results from existing numerical methods demonstrates that the surrogate model can predict responses for various geometries of tools, soil parameters, and moving paths.The second part of the work focuses on evaluating the effectiveness and efficiency of physics-informed neural networks (PINNs) in inverse analysis for soil plasticity. Compared to inverse analysis with surrogate models built through black-box machine learning, PINNs have the advantage of requiring significantly less data. A PINN is developed specifically for elastoplastic materials. To capture the history-dependent properties of such materials, a recurrent neural network (RNN) is used. Before engaging in inverse analysis, the proposed model is validated in forward analysis through two typical tests for soils, i.e., biaxial compression and flat punch indentation. Results from the model closely align with finite element solutions. In terms of inverse analysis, the model successfully identifies plastic parameters of elastoplastic materials without the need for iterative simulations. To eliminate some efforts in manually adjusting the loss function of the neural network, a hard imposition of measurements is employed, automatically ensuring that the outputs of the neural network adhere to the measured data. To accelerate the inverse analysis process, different strategies of transfer learning are also explored.
546 $a English
590 $a School code: 0163
650 4 $a Geotechnology. $3 1018558
650 4 $a Engineering. $3 586835
650 4 $a Computer science. $3 523869
650 4 $a Environmental engineering. $3 548583
653 $a Inverse analysis
653 $a Large deformation
653 $a Machine learning
653 $a Physics-informed neural networks
653 $a Plasticity
653 $a Surrogate model
690 $a 0428
690 $a 0984
690 $a 0775
690 $a 0537
710 2 $a Northwestern University. $b Civil and Environmental Engineering. $e degree granting institution. $3 3815791
720 1 $a Hambleton, James $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 85-10B.
790 $a 0163
791 $a Ph.D.
792 $a 2024
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30993550