% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{RajaeiHarandi:1017335,
author = {Rajaei Harandi, Ali},
othercontributors = {Reese, Stefanie and Wessels, Henning},
title = {{B}ridging classical and deep learning approaches for
multiscale and multiphysics systems},
school = {Rheinisch-Westfälische Technische Hochschule Aachen},
type = {Dissertation},
address = {Aachen},
publisher = {RWTH Aachen University},
reportid = {RWTH-2025-07322},
pages = {1 Online-Ressource : Illustrationen},
year = {2025},
note = {Veröffentlicht auf dem Publikationsserver der RWTH Aachen
University; Dissertation, Rheinisch-Westfälische Technische
Hochschule Aachen, 2025},
abstract = {Advances in the production of engineered materials and the
tools used to produce them are widening the scope for
achieving ambitious design goals and developing
sophisticated materials with exceptional performance
characteristics. The successful attainment of desired
mechanical properties in advanced materials is fundamentally
dependent on the deliberate design and control of their
microstructure. The behavior observed at the macroscale -
such as intrinsic or process-induced anisotropy, plastic
deformation, and damage evolution - can all be traced back
to the underlying architecture at smaller scales. Whether it
is grain orientation in polycrystals, fiber distribution in
composites, or columnar grain morphology in hard coatings,
microstructural features serve as the governing framework
for the overall materials’ response. While the
experimental study of microstructure evolution and material
behavior remains acritical part of materials science, it is
often time-consuming, labor-intensive, and expensive.
Increasing computational power, the development of advanced
numerical methods, and the emergence of machine learning
techniques have significantly expanded the flexibility and
efficiency of materials design processes. Modeling complex
material behavior often involves coupled phenomena such as
plasticity, damage, phase transformation, and anisotropy,
and requires extending traditional continuum models to
capture these multiphysical interactions. This, in turn, not
only increases the computational cost of simulations, but
also introduces several numerical challenges-such as
stability issues, convergence difficulties, and
discretization sensitivity that are often exacerbated by the
inherent complexity of these models. As a result, such
simulations can be error-prone and require careful
calibration and validation to ensure reliability. In this
context, deep learning approaches have emerged as powerful
surrogates capable of accelerating both local (e.g.
constitutive material models) and global (e.g. structural
response) simulations, enabling faster and more scalable
predictions without compromising accuracy. This cumulative
dissertation aims to develop efficient numerical and deep
learning-based models for accurately predicting the
mechanical behavior of heterogeneous materials. The overall
objective is to build a generic framework for modeling
heterogeneous microstructure and show the connection of
several material properties to the microstructure. The
ultimate goal is to accelerate the material design process
while ensuring their durability and performance in
accordance with specific application requirements. This
compilation of scientific papers by the author and
co-authors presents advanced modeling techniques that help
design complex materials to support more sustainable
development. The first two papers use an anisotropic
cohesive phase field approach to study the mechanical
behavior of hard coatings with heterogeneous
microstructures, fine columnar grain morphologies. In the
first paper, the developed computational framework is
implemented to analyze the influence of several key
parameters, such as residual stress, crack initiation
stress, and grain morphology, on the cracking behavior of
hard coatings subjected to micro-tensile loading. The model
incorporates microstructure-informed fracture energy to
accurately represent the fracture behavior in these
heterogeneous coatings as well as to account for the proper
softening behavior (cohesive-like behavior) at the
micro-scale. A comparative study is conducted against
experimental data obtained from micro-tensile tests. Both
qualitative and quantitative comparisons demonstrate the
predictive capability of the proposed methodology in
capturing crack initiation and propagation in VAlN coatings
deposited by high-power pulsed magnetron sputtering (HPPMS).
In the second work, the framework is extended to simulate
crack initiation and evolution under compressive loading
conditions. The anisotropic cohesive phase-field model is
coupled with a fracture-motivated driving force, which
accounts for the energy contributions from principal stress
components in the damage driving force. This allows the
model to capture crack initiation stresses and fracture
energies associated with different fracture modes beyond the
pure tensile opening. The methodology is applied to simulate
the fracture behavior of an isotropic hard coating layers
subjected to micro-pillar compression tests. The third paper
presents a comparative analysis between standard phase-field
models and gradient-extended damage models, both from a
theoretical perspective and in terms of parameter
correspondence. The investigation focuses on establishing a
connection between the governing parameters of the two
frameworks to ensure they yield consistent predictions under
equivalent conditions. In pursuit of a computationally
efficient surrogate modeling strategy, the fourth work
intro-duces a mixed physics-informed neural network (PINN)
framework tailored for thermoelastic problems. To
effectively capture the influence of material heterogeneity,
the proposed model employs separate neural networks to
approximate the primary field variables (such as temperature
and displacement) and their associated spatial gradients
(stresses and heat fluxes). The training process leverages
both coupled and sequential strategies. In the sequential
approach, the network parameters corresponding to one
physical domain (either thermal or mechanical) are frozen
while optimizing the loss function for the other, allowing
for more stable and ac-curate convergence. Furthermore, by
incorporating heterogeneity maps as additional input
features, the methodology is extended to generalize across a
wide range of material property combinations, enabling the
model to handle varying ratios of thermal and mechanical
material parameters. In the fifth paper, a novel
physics-informed operator learning strategy is introduced.
The neural operator is trained to map a wide range of
microstructures to their corresponding local stress fields.
This is achieved by incorporating a fixed-point iteration
scheme within FFT-based frameworks, thereby eliminating the
need for automatic differentiation to formulate the
underlying partial differential equations. By working at a
fixed resolution, the loss function is constructed using the
Lippmann–Schwinger operator in Fourier space. The proposed
model demonstrates excellent scalability, effectively
generalizing to previously unseen microstructures. On the
one hand, the loss formulation significantly reduces
training time. On the other, the Fourier Neural Operator
architecture exhibits strong performance in predicting
material responses.},
cin = {311510},
ddc = {624},
cid = {$I:(DE-82)311510_20140620$},
pnm = {DFG project G:(GEPRIS)259792543 - Mehrskalige Modellierung
des Schädigungs- und Bruchverhaltens nanostrukturierter
Schichten (A06) (259792543) / TRR 87: Gepulste
Hochleistungsplasmen zur Synthese nanostrukturierter
Funktionsschichten},
pid = {G:(GEPRIS)259792543 / G:(GEPRIS)138690629},
typ = {PUB:(DE-HGF)11},
doi = {10.18154/RWTH-2025-07322},
url = {https://publications.rwth-aachen.de/record/1017335},
}
guest ::