E-Thesis 440 views
The data-driven-based constitutive model and its application in machine learning-aided multiscale modelling for granular materials / MENGQI WANG
Swansea University Author: MENGQI WANG
DOI (Published version): 10.23889/SUThesis.69879
Abstract
Data-driven methods have gained significant attention since Google’s AlphaGo defeated a world champion in 2017. Over the past several years, machine learning (ML), a subset of data-driven approaches, has increasingly attracted interest in the study of granular materials within civil engineering. Unli...
| Published: |
Swansea University, Wales, UK
2025
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Feng, Y. T. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa69879 |
| Abstract: |
Data-driven methods have gained significant attention since Google’s AlphaGo defeated a world champion in 2017. Over the past several years, machine learning (ML), a subset of data-driven approaches, has increasingly attracted interest in the study of granular materials within civil engineering. Unlike traditional techniques, such as phenomenological models and numerical methods used to describe the deformation of granular media, the ML method offers several distinct advantages: 1) ML models can continuously adapt to new data, enhancing their accuracy and performance over time, thereby demonstrating strong adaptability; 2)ML techniques can directly uncovering hidden relationships and extracting insights from data without any assumption, making them particularly effective for predictive analytics;3) ML algorithms execute complex calculations at high speeds, significantly reducing the time required for various tasks. These features of the ML methods make them a promising alternative to bypass the challenges inherent in traditional research methodologies, such as complex assumptions, limited adaptability, and high computational costs. This thesis aims to develop a data-driven multiscale modelling method incorporating an ML-based stress-strain model, capable of precisely capturing history-dependent constitutive behaviours for grain assembly, to accelerate the computational process of the traditional FEM-DEM approach in both 2D and 3D simulations. The whole work comprises the following three main chapters:1. Literature review for machine learning aided modelling of granular materials: The recent advances in ML-aided studies of granular materials are reviewed. Following the literature review work, the key challenges in developing a DEM data-driven FEM-ML framework are identified and discussed.2. The temporal convolution neural network (TCNN)-based constitutive model: The potential of the TCNN-based constitutive model in capturing the mechanical behaviours of granular materials is comprehensively investigated and the applicability of time-sequence ML models in developing the FEM-ML framework is discussed3. The enhanced multi-layer perception (MLP)-based FEM-ML modelling algorithm: A DEM data-based FEM-ML framework is constructed with the enhanced MLP that integrates Frobenius norm-based internal variable. The data-driven multiscale algorithm significantly improves the solving efficiency of the traditional FEM-DEM method in 2D BVPs for granular material.4. A physical information-integrated FEM-ML multiscale framework for 3D modelling of granular materials: Three internal variables with physical significance are introduced to develop the MLP-based FEM-ML modelling method for 3D problems.The modelling results obtained by the FEM-ML method are comparable to those from the FEM-DEM approach but achieve significant improvement in computational efficiency. |
|---|---|
| Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information. |
| Keywords: |
Granular materials, Machine learning, Constitutive model, FEM, DEM, FEM-DEM multiscale method, FEM-ML method |
| College: |
Faculty of Science and Engineering |
| Funders: |
CSC |