|
Times |
MF 2:30-3:20PM, Phillips 385 |
Office hours |
MF 12:30-2:00PM, and by email appointment, Chapman 451 |
Instructor |
This directed reading course investigates the theory of continuum mechanics from both a classical and contemporary perspective. Classical continuum mechanics is usually presented using the tools of differential calculus and provides a complete description for linear media with no memory effects, as exemplified most prominently by the Cauchy elasticity equations. While sufficient for the purposes of traditional mechanical engineering relying on small deformation of crystalline metals, vast classes of important materials encountered in plastics processing, paper processing, non-Newtonian flow or biological materials are only awkwardly described through the mathematical apparatus of partial differential equations (PDEs) mainly due to four causes:
Many materials of practical contemporary interest exhibit intrinsic nonlinear behavior, typically arising from a mesoscopic underlying structure instead of a microscopic one. Whereas the mechanical behavior of metallic monocrystals arises from electrostatic interaction acting at the length scale of the crystalline lattice of m along all directions, cytoskeleta exhibit an underlying structure with an average actin filament length of m with a radius of m, while muscle tissue fibers are of length m with a radius of m. When averaging out, or homogenizing, the behavior of the constituent elements of the medium to a macroscopic length scale of practical interest, say m, the large number of interactions in materials with microscopic structure often leads to isotropic and linear partial differential equations, but anisotropic and nonlinear behavior is obtained for materials with mesoscopic structure.
In classical continuum mechanics the underlying microscopic structure is typically considered to be fixed in time, a good approximation of the slow rate of chemical reactions (e.g., oxidation of iron) within the range of materials of interest. In contrast, conformational changes in flowing polymers or cellular dephosphorylation of ATP into ADP leads to markedly different mechanical behavior in viscoelastic flow or cellular motility. Such materials are said to be active, and typically exhibit very complex and incomplete mathematical descriptions within the framework of differential equation theory.
Active materials tend to rearrange their mesoscopic structure, typically under the influence of random excitation (i.e., thermal bath) from the surrounding medium. This leads to the need for stochastic processes to describe the response of the medium to application of external forces.
The reorganization of mesoscopic structures due to chemical activity or external forces often occurs at time scales longer than those of observation, such that the prior history of the medium influences the observed behavior. Such memory effects can be modeled by differential equations of fractional order (i.e., integro-differential equations), and when combined with random thermal forcing require the consideration of non-Markovian stochastic processes.
The classical approach to address the above features of complex materials has been to posit algebraic or differential constitutive relations between deformation and stress, and investigate the validity of the hypothesized models. An alternative suggested by current progress in machine learning is to extract the constitutive relations directly from data.
Tensor algebra and calculus
Kinematics of point masses, rigid bodies, deformable bodies
Conservation laws (mass, momentum, energy) applied to continua
Elasticity, reversible small-amplitude deformations of a solid
Plasticity, irreversible deformation of a solid
Navier-Stokes flow of a Newtonian fluid
Viscoelastic flow
Fiber networks
Active media
Review of least squares approximation, seen as a particular example of machine learning
Review of artificial neural networks, seen as a particular example of approximation theory
Deep neural network models of constitutive equations:
Classical elasticity theory recovered from a deep neural model of linear deformation
Deep neural model of plastic deformation
Navier-Stokes equations recovered from deep neural model of Newtonian flow
Deep neural models for viscoelastic flow
Deep neural models of fiber networks
Stochastic neural network models of active media
Polymeric fluids with changing connectivity, reptation.
Fiber networks with changing connectivity
Fiber networks with active elements
There is no single course text. Topics are drawn from the following sources that are available in electronic form from the course repository or UNC library.
Machine Learning and Continuum Mechanics, S. Mitran
Continuum Mechanics: Foundations and Applications of Mechanics, C.S. Jog
Continuum Mechanics, Andrus Koppel and Jaak Oja
Elements of continuum mechanics, R.C. Batra
Deep Learning, I. Goodfellow and Y. Bengio and A. Courville
Additional material is available in the course material repository (/biblio subdirectory)
Lessons |
Topics |
TEN, KIN, CON |
|
ELS |
|
DNN-ELS |
|
PLS |
|
DNN-PLS |
|
NSF |
|
DNN-NSF |
|
VEF |
|
DNN-VEF |
|
NET |
|
DNN-NET |
|
SNN-NET |
|
ACT |
|
SNN-ACT |
The following codes generate data sets for various continuum modles
Nr. |
Topic |
Problems |
Solutions |
1 |
LSQ, TensorFlow |
|
|
2 |
ELS |
|
|
3 |
PLS |
|
|
4 |
NSF |
|
|
5 |
VEF |
|
|
6 |
NET |
|
|
7 |
ACT |
|
Though the course concentrates on concepts within continuum mechanics and machine learning, the computational implementation of these concepts is essential to an appreciation of the utility of the considered approaches. Templates are provided for all computational applications, typically comprising:
Generation of data for constitutive relations;
Definition of deep neural networks that approximate the constitutive relation data;
Numerical methods (finite element, finite volume) that solve the classical formulations of continuum mechanics, e.g., Cauchy elasticity equations, Navier-Stokes flow equations.
This course uses a customized Linux environment named SciComp@UNC available to students as a virtual machine in which all course software is preinstalled, and course applications are preconfigured. Download Virtual Box and the SciComp@UNC virtual machine image.
Various software tools for carrying out and documenting practical scientific computation will be successively introduced:
TeXmacs: editing of documents containing live computation
SciPy: scientific Python environment
TensorFlow: machine learning platform accessible from Python
Mathematica: system for numerical, symbolic, and graphical computation with DNN support
Gnu compilers (Fortran, C++): high-performance compiled code development
Julia: a high-performance interactive environment
Paraview: data visualization
BEARCLAW: a package for solving PDEs using finite volumes
FreeFEM++: a package for solving PDEs using finite elements
The Mathematica commercial package is accessible to students while connected to the campus network (either directly or remotely through the UNC VPN server).
Course materials (lecture notes, workbooks, homework, examination examples) are stored in a repository that is accessed through the subversion utility, available on all major operating systems. The URL of the material is http://mitran-lab.amath.unc.edu/courses/MATH768.
The initial svn checkout is made using commands:
mkdir ~/courses cd ~/courses svn co svn://mitran-lab.amath.unc.edu/courses/MATH768
On SciComp@UNC the initial checkout can be carried out through the terminal commands:
cd ~/courses make MATH768
Update the course materials before each lecture by:
cd ~/courses/MATH768 svn update
Links to course materials will also be posted to this site, but the most up-to-date version is that from the subversion repository, so carry out the svn update procedure prior to each lecture.