12 March (Fri), 2021
10:20–17:30 (JST), 9:20–16:30 (Taiwan Time)
Online by Zoom. Click here to join the workshop.
(Meeting ID : *************, Passcord : ******)
Note that they are subject to change.
Zoom connection code will be also informed via your registered e-mail address.
The following venue is also available:
Room 203 (Applied Analysis Seminar Room) at
Integrated Research Bldg. No.12([54] in map), Kyoto University.
- Please change your display name in Zoom to "Your Name (Your Affiliation)".
- Recoding (video/audio) is not allowed.
- You are recommened to register your participation with Google Form.
Hau-Tieng Wu (Duke University)
Yi-Hsuan Lin (National Chiao-Tung University)
Pu-Zhao Kow (National Taiwan University)
Pengwen Chen (National Chung-Hsing University)
Chin-Tien Wu (National Chiao-Tung University)
Michael Koch (Kyoto University) Michael Koch (京都大学)
Yu-Hsun Lee (Kyoto University) 李昱勳 (京都大学)
Kazunori Matsui (Kanazawa University) 松井一徳 (金沢大学)
Takaaki Nishida (Kyoto University) 西田孝明 (京都大学)
Masahiro Yamamoto (The University of Tokyo) 山本昌宏 (東京大学)
${}^1$Graduate School of Agriculture, Kyoto University
Kitashirakawa Oiwake-cho, Sakyo-ku, 606-8502, Kyoto (Japan)
Abstract: We consider the simultaneous identification of the interface of an anomaly/cavity (embedded in a domain $\Omega$) along with the Gaussian spatial random field $k(\mathbf{z}, \omega)$ defined at each point $\mathbf{z} \in \Omega$, from noisy observation data. Here, $\omega$ is an element of the sample space $\Theta$, and $(\Theta, F, P)$ is a complete probability space. Computational solution of the problem necessitates the discretization of the random field, which is conveniently done through the $K$-term truncated Karhunen-Loève (KL) expansion, \begin{equation} k(\mathbf{z},\omega)\approx \overline{k}(\mathbf{z}) + \sum_{i=1}^K \sqrt{\lambda_i}\varphi_i(\mathbf{z})^1\theta_i(\omega), \end{equation} where $\overline{k}(\mathbf{z})$ is the mean function and ${ }^1\theta_i(\omega)$ are standard normal random variables. $\lambda_i$ and $\varphi_i(\mathbf{z})$ are obtained from the solution of the Integral Eigen Value Problem (IEVP) \begin{equation} \int_\Omega C(\mathbf{z},\mathbf{z}')\varphi_i(\mathbf{z}')d\mathbf{z}' = \lambda_i\varphi_i(\mathbf{z}), \end{equation} where $C(\mathbf{z}, \mathbf{z}')$ is the autocovariance function, which is symmetric and positives semi-definite. The eigen values $\lambda_i$ decay rapidly and for most practical purposes, the KL expansion involves a summation over a few terms only. Except for a few simple problems, the IEVP has to be solved numerically over $\Omega$.
Statistical inversion is carried out in the Bayesian framework through a modern gradient based Markov Chain Monte Carlo (MCMC) algorithm called Hamiltonian Monte Carlo or HMC that avoids random walks by making gradient guided proposals. The parameter space is composed of two kinds of parameters ${ }^1\boldsymbol{\theta}$ (representing the spatial field) and ${ }^2\boldsymbol{\theta}$ (representing the interface). As the interface is updated at every step of HMC, the domain $\Omega$ is also updated and the IEVP has to be solved again. Computational complexity is further increased as the HMC gradients require the computation of the gradients $\dfrac{\partial\lambda_i}{\partial{ }^j\boldsymbol{\theta}}$ and $\dfrac{\partial\varphi_i}{\partial{ }^j\boldsymbol{\theta}}$ , where $j = 1,2$.
For efficient computation, we leverage the domain independence property of the KL expansion, which states that the first and second order moments of a random process generated by the KL expansion are invariant to a change in the physical domain. Hence, $\lambda_i$ and $\varphi_i(\mathbf{z})$ are computed only once using the highly efficient Nyström method on a rectangular bounding domain $\Omega_b$, such that every realization in HMC of $\Omega \subseteq \Omega_b$. The eigen vectors are then interpolated to the FE mesh Gauss points at every step of HMC. The domain independence property further enables computational savings as the gradients of the eigen values and eigen vectors now no longer have to be calculated. This method is employed to determine the hydraulic conductivity $k(\mathbf{z},\omega)$ and the length and width of a piping zone embedded in a seepage zone shown in Fig 1, using experimental data.
Fig. 1(a) Rectangular bounding domain $\Omega_b$ that bounds all HMC realizations $\Omega$. Black circles represent hydraulic head observation points in the seepage zone. (b) $20\times20$ grid of Gauss-Legendre quadrature points to solve the IEVP using the Nyström method on $\Omega_b$. (c) Domain independence of the KL expansion demonstrated through the probability density functions at the FEM Gauss point (marked as blue circle in (a)). Here, the IEVP is solved through Galerkin projection, on domain $\Omega$ and through the Nyström method on $\Omega_b$.
The conventional fractional derivative of order $\alpha < 1$ requires the first-order derivative and so we need some justification especially when we consider a weak solution. Another theoretical issue is the formulation of the initial condition because for $\alpha<\frac{1}{2}$ we cannot expect convenient continuity at $t=0$.
First, in order to solving these issues, we define $\partial_t^{\alpha}$ in the subspaces $H_{\alpha}(0,T)$ in usual Sobolev-Slobodeckij spaces $H^{\alpha}(0,T)$, which is the clousure operator of the Caputo derivative in $\{w\in C^1[0,T];\, w(0) = 0\}$. Then we establish fundamental properties such as isomorphism between $\Vert \partial_t^{\alpha} w\Vert_{L^2(0,T)}$ and $\Vert w\Vert _{H_{\alpha}(0,T)}$, and we prove the unique existence of the solution to the initial boundary value problem with regularity properties. As for the details, we refer to Kubica, Ryszewska and Yamamoto [1].
Second, based on such a theory, I present the following topics among recent theoretical results for several inverse problems:
- Asymptotic behavior of the solution as $t\to \infty$
- Backward problems in time
- Determination of orders $\alpha$
- Inverse coefficient problems
Especially I would like to stress that fractional differential equations are promissing to the young, because this field is strongly supported by many real-world problems and applications such as anomalous diffusion of contaminants, and reliable mathematical researches are highly demanded.
References
[1] A. Kubica, K. Ryszewska, and M. Yamamoto,
Time-fractional Differential Equations:
a Theoretical Introduction, Springer Japan, Tokyo, 2020.
Abstract:
In order to obtain the roll-type solution of Rayleigh-Bénard heat convection for large Rayleigh number we use the stream function and temperature as unknowns : \begin{align} \dfrac{\partial}{\partial t}\Delta\psi &= \mathcal{P}_r\Delta^2\psi + \mathcal{P}_r\mathcal{R}_a\dfrac{\partial\theta}{\partial x} + \dfrac{\partial\psi}{\partial z}\dfrac{\partial\Delta\psi}{\partial x} - \dfrac{\partial\psi}{\partial x}\dfrac{\partial\Delta\psi}{\partial z}\tag{$1$}\\ \dfrac{\partial\theta}{\partial t} &= \Delta\theta + \dfrac{\partial\psi}{\partial x} + \dfrac{\partial\psi}{\partial z}\dfrac{\partial\theta}{\partial x} - \dfrac{\partial\psi}{\partial x}\dfrac{\partial\theta}{\partial z}.\tag{$2$} \end{align} Since the boundary conditions are stress free, the solution has the Fourier series expansion. We express $(1)$ $(2)$ with Fourier series expansions and use Galerkin method and Newton method for the solutions of the stationary system of $(1)$ $(2)$.
$\mathcal{R}_a = 20.0\times\mathcal{R}_c$, $\mathcal{P}_r = 10.0$
$N$ | $\mathcal{R}_a/\pi^4$ | $r=\mathcal{R}_a/\mathcal{R}_c$ | $\lambda$ |
$32$ | $277.76301\cdots$ | $41.15007\cdots$ | $i\times40.75766\cdots$ |
$40$ | $277.40075\cdots$ | $41.09640\cdots$ | $i\times40.72751\cdots$ |
$48$ | $277.36731\cdots$ | $41.09145\cdots$ | $i\times40.72462\cdots$ |
$56$ | $277.36495\cdots$ | $41.09110\cdots$ | $i\times40.72441\cdots$ |
$64$ | $277.36481\cdots$ | $41.09108\cdots$ | $i\times40.72440\cdots$ |
(Step 1) Find $u^*_k:\Omega\rightarrow\mathbb{R}^d$; \begin{align*}\left\{\begin{array}{ll} {\displaystyle \frac{u^*_k-u_{k-1}}{\tau}+D(u^*_{k-1},u^*_k) -\nu\Delta u^*_k=f(t_k)}&\mbox{in }\Omega,\\[8pt] u^*_k=0&\mbox{on }\Gamma_1,\\ u^*_k\times n=0&\mbox{on }\Gamma_2,\\ \text{div}\, u^*_k=0&\mbox{on }\Gamma_2. \end{array}\right.\end{align*}
(Step 2) Find $P_k:\rightarrow\mathbb{R}, u_k:\rightarrow\mathbb{R}^d$; \begin{align*}\left\{\begin{array}{ll} {\displaystyle -(\tau/\rho)\Delta P_k=-\text{div}\,u^*_k} &\mbox{in }\Omega,\\[4pt] {\displaystyle \frac{\partial P_k}{\partial n}=0} &\mbox{on }\Gamma_1,\\[8pt] P_k=p^b(t_k)&\mbox{on }\Gamma_2, \end{array}\right.\\ {\displaystyle u_k=u^*_k-(\tau/\rho)\nabla P_k}\quad\mbox{in }\Omega.\quad \end{align*}
We show a stability of the scheme and establish error estimates in suitable norms between the solutions to the scheme and (NS).
Jenn-Nan Wang (National Taiwan University) 王振男 (臺灣大學)
Yuusuke Iso (Kyoto University) 磯祐介 (京都大学)
Hiroshi Fujiwara (Kyoto University) 藤原宏志 (京都大学)
Hitoshi Imai (Doshisha University) 今井仁司 (同志社大学)
Local Organizers
Daisuke Kawagoe (Kyoto University) 川越大輔 (京都大学)
Gi-Ren Liu (National Cheng Kung University) 劉聚仁 (國立成功大學)