Multi-field and multi-scale computational fracture mechanics and machine-learning material modeling

• Mehrfeld- und mehrskalige rechnerunterstützte Bruchmechanik und maschinell lernende Materialmodellierung

Heider, Yousef; Markert, Bernd (Thesis advisor); Schrefler, Bernhard (Thesis advisor); Müller, Ralf (Thesis advisor)

Aachen : RWTH Aachen University (2021)
Book, Habil / Postdoctoral Thesis (Non-german Habil)

In: Report. IAM, Institute of General Mechanics IAM-13
Page(s)/Article-Nr.: 1 Online-Ressource : Illustrationen, Diagramme

Habilitationsschrift, Rheinisch-Westfälische Technische Hochschule Aachen, 2021

Abstract

Fracture mechanics counts to the most emerging and promising fields of engineering mechanics. In the last few decades, the topics of crack initiation and propagation in solid and porous materials have attracted numerous theoretical, experimental, and numerical studies. This was driven by many challenges and necessities in engineering fields, such as the bad need for designing safe, reliable, and sustainable structures that withstand all types of expected natural and human actions, or the promising application of fracture tools in sectors like energy production, geothermal systems, soil science, and geotechnical engineering. From a mechanical and computational point of view, the fracturing of solid and porous materials presents a challenging multi-scale multi-phase problem, which includes possible several simultaneous physical processes and many sources of numerical instability. For a holistic understanding as well as efficient and accurate fracture modeling, the underlying monograph will address fracture mechanics and related processes across the scales, i.e. nanoscale, microscale, and macroscale. This includes, first, utilization of Molecular Dynamics (MD) simulations to understand fracture mechanism and conclude material parameters of brittle solid materials on the nanoscale, second, embedding the phase-field modeling (PFM) approach in continuum mechanics for fracture modeling on the macroscopic scale, and, third, embedding the PFM approach in continuum porous media mechanics (PM) to model hydraulic fracturing in saturated and unsaturated porous media, i.e. PM-PFM combined procedure. In conventional approaches in the mechanics of materials, such as in fracture mechanics, solid mechanics, or porous media mechanics, the constitutive modeling provides explicit mathematical expressions, which are based on phenomenological observations or experimental data. These models can further be subjected to hard constraints, such as the balance equations or the thermodynamics restrictions. To avoid the constitutive model's complexity and the increase of the number of required material parameters to an impractical level, these material models partially or entirely overlook microscopic information. This might lead, however, to deterioration of the model's accuracy, especially in the description of multi-scale and time- or path-dependent responses like in crystal plasticity or in nonlinear anisotropic porous media flow. This paves the way for the implementation of data-based artificial neural networks (ANN) to generate machine-learning (ML)-material models, which are capable to extract complex dependencies on micro-geometry and time or path dependencies without the need to explicitly determine the material parameters. Therefore, the fourth aim of the underlying monograph will be utilizing the capabilities of Machine Learning, via using deep neural networks (DNN) and deep reinforcement learning (DRL) to generate ML-based material models, which rely on microstructural information in the training datasets. The aforementioned approaches backed by powerful computational capacities and experimental data give the ability to reliably simulate and understand complicated real multi-phase and multi-scale problems out of solid and porous media mechanics.