Data driven constitutive models for engineering materials
File(s)
Author(s)
Ge, Weijian
Type
Thesis
Abstract
The primary objective of this PhD thesis is to develop data-driven models for predicting the behavior of engineering materials. By doing so, we aim to enhance the efficiency of multiscale simulations and transition from empirical and arbitrary approaches to more reliable engineering predictions. This work offers an overview of computational homogenization principles and the latest applications of machine learning in solid mechanics.
Initially, we establish a computational framework for n-phase composites, incorporating hyperelastic, viscoelastic, and plasticity behaviours. We introduce a random walk algorithm for generating multi-axial loading paths to compile training data and employ a neural network to learn the macroscopic behaviour. We then evaluate the surrogate models' accuracy and investigate how their precision is affected by the composites' heterogeneity and the training dataset's size. The findings reveal that the proposed models effectively capture the non-linear multi-axial material response across extensive strain rates and strain triaxiality, including non-monotonic and non-proportional loading.
Subsequently, we expand the framework to encompass 6D loading cases, focusing on progressive damage in ductile isotropic and anisotropic elastic damage situations. We devise an effective elastic material and linear perturbation to assess overall damage in both instances. In the ductile damage case, we employ an innovative training method to identify damage initiation. For anisotropic damage cases, we utilize dimensional reduction techniques to estimate the overall damage parameter. The data-driven models developed for both scenarios accurately represent the loss of stiffness and the overall stress-strain curve. These models exhibit a robust capacity to capture the material's progressive damage response, resulting in highly precise predictions of material behaviour under arbitrary loading conditions...
Initially, we establish a computational framework for n-phase composites, incorporating hyperelastic, viscoelastic, and plasticity behaviours. We introduce a random walk algorithm for generating multi-axial loading paths to compile training data and employ a neural network to learn the macroscopic behaviour. We then evaluate the surrogate models' accuracy and investigate how their precision is affected by the composites' heterogeneity and the training dataset's size. The findings reveal that the proposed models effectively capture the non-linear multi-axial material response across extensive strain rates and strain triaxiality, including non-monotonic and non-proportional loading.
Subsequently, we expand the framework to encompass 6D loading cases, focusing on progressive damage in ductile isotropic and anisotropic elastic damage situations. We devise an effective elastic material and linear perturbation to assess overall damage in both instances. In the ductile damage case, we employ an innovative training method to identify damage initiation. For anisotropic damage cases, we utilize dimensional reduction techniques to estimate the overall damage parameter. The data-driven models developed for both scenarios accurately represent the loss of stiffness and the overall stress-strain curve. These models exhibit a robust capacity to capture the material's progressive damage response, resulting in highly precise predictions of material behaviour under arbitrary loading conditions...
Version
Open Access
Date Issued
2023-03-16
Date Awarded
2024-01-01
License URL
Advisor
vito tagarielli, Vito
Publisher Department
Aeronautics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)