Research

PhD topic

The project's title was Adaptive numerical algorithms for PDE problems with random data and my supervisor has been Alex Bespalov.

The project concerned the design of efficient adaptive algorithms for the numerical approximation of solutions to elliptic Partial Differential Equations (PDEs) with parametric or random inputs. The underlying numerical method used was the Finite Element Method (FEM).

In many engineering applications, we have to account for quantities such as diffusion coefficients, boundary, and source terms which are affected by uncertainty, either because of an incomplete knowledge (epistemic uncertainty) or due to an intrinsic variability of the physical system under consideration (aleatoric uncertainty).

In a typical PDE-based setting, it is difficult to accurately describe the model and, in most of cases, a fully deterministic description of it cannot be obtained. Therefore, there is the need of reliable uncertainty quantification. This is achieved by supplying the model with a probabilistic framework (i.e., by introducing random variables, random fields, etc.).

Model problem

The model problem we are interested with is the following one.

Given a source term $f \in L^2(D)$, we consider the solution $u:D\times \Gamma\to \mathbb{R}$ to the following parametric steady-state diffusion equation with homogeneous Dirichlet boundary conditions, \[ \begin{align} -\nabla \cdot (a(\textbf{x},\textbf{y})\nabla u(\textbf{x},\textbf{y})) &= f(\textbf{x}) && \textbf{x} \in D, \, \textbf{y} \in \Gamma, \\ u(\textbf{x},\textbf{y}) &= 0 && \textbf{x} \in \partial D, \, \textbf{y} \in \Gamma, \end{align} \] where $D \subset \mathbb{R^2}$ is a bounded polygonal spatial domain, $\Gamma$ is the parameter domain, defined as $\Gamma := \bigotimes_{m\geq 1} \Gamma_m$ with $\Gamma_m := [-1,1]$, $\textbf{y} = (y_1,y_2,\dots) \in \Gamma$ is a (multivariate) random parameter, and finally, $a(\textbf{x},\textbf{y})$ is a second-order random field represented as \[ a(\textbf{x},\textbf{y}) = a_0(\textbf{x}) + \sum_{m\geq 1} a_m(\textbf{x})\, y_m, \quad \textbf{x} \in D, \; \textbf{y} \in \Gamma, \] for a given family of regular functions $a_m(\mathbf{x}) \in L^{\infty}(D)$, for all $m \in \mathbb{N}_0$.

For example, $y_m \in \Gamma_m$ can be defined as the images of independent uniformly distributed mean-zero random variables on $\Gamma_m$, for all $m \in \mathbb{N}$.

Method

The method used to solve our problem is the Stochastic Galerkin Finite Element Method (SGFEM). This method couples together the degrees of freedom from the discretisation of the spatial and the parameter domain and it generates solutions in tensor product spaces $X \otimes \mathcal{P}$, where $X$ is a finite element space and $\mathcal{P}$ is a space of (multivariate) polynomials in the parameters.

Unlike more classic sampling based methods, such as Monte Carlo and Stochastic Collocation methods, the SGFEM solves just one (very) large linear system instead of a (very) large number of smaller linear systems arising from the discretisation associated with each fixed sample $\mathbf{y} \in \Gamma$.

Thesis

You can download my PhD thesis here (or alternatively here).

General Documents

See here for a set of general documents, lectures, and spare notes I decided to save during the PhD.

Experience

During my doctoral studies I have been working as teaching assistant for the following undergraduate modules at the School of Mathematics of the University of Birmingham:
  • Computer Lab Demonstrating – Spring 2016, Spring 2017
  • Methods in Partial Differential Equations – Spring 2016
  • 1st-year Support Class – Spring 2016, Autumn 2016
  • Numerical Methods II – Autumn 2016, Autumn 2017, Autumn 2018
  • Linear Programming – Autumn 2016, Autumn 2017, Autumn 2018
  • Mathematical Modelling and Problem Solving – Autumn 2016
  • Mathematics in Industry – Spring 2017
  • Differential Equations – Spring 2017
  • Numerical Linear Algebra with Applications – Spring 2018
  • Numerical Methods and Programming – Spring 2018

The main duties were:

  • Assisting and tutoring the undergraduate students;
  • Helping them with coding exercises in laboratory classes;
  • Invigilating tests and exams, as well as marking continuous assessments.

Events

  • 5–6 December, 2016 – Attending the workshop Adaptive algorithms for computational PDEs at the University of Birmingham.
  • 14–17 June, 2016 – Attending the conference MAFELAP: the Mathematics of Finite Element and Applications, Brunel University, London.
  • 12–16 September, 2016 – Attending the summer school Uncertainty quantification, Weierstrass Institute (WIAS), Berlin.
  • 6 March, 2017 – Giving a talk at the postgraduate seminars (LAMS) of the School of Mathematics of the University of Birmingham. Talk's title: Adaptive Finite Element algorithms for PDEs with random data.
  • 27–30 June, 2017 – Giving a talk at The 27th Biennial Numerical Analysis Conference, University of Strathclyde, Glasgow. Talk's title: An adaptive stochastic Galerkin FEM for parametric PDEs with singular solutions; slides.
  • 31 January 2018 – Taking part in a student poster competition held at the School of Mathematics of the University of Birmingham; poster.
  • 9 March 2018 – Giving a talk at the postgraduate seminars (LAMS) of the School of Mathematics of the University of Birmingham. Talk's title: Efficient adaptive algorithms for PDEs with random input data; slides.
  • 19 June 2018 – Giving a talk at the SIAM UKIE National Student Chapter Conference held at the University of Bath. Talk's title: An efficient adaptive algorithm for elliptic problems with random input data; slides.

Publications

  • A. Bespalov, L. R., and D. Silvester, T-IFISS: a toolbox for adaptive FEM computation, Comput. Math. Appl., vol 81, pp. 373-390, Jan 2021; download (Elsevier page).
  • A. Bespalov, D. Praetorius, L. R., and M. Ruggeri, Convergence of adaptive stochastic Galerkin FEM, SIAM J. Numer. Anal., vol. 57, no. 5, pp. 2359-2382, 2019; download (SIAM page).
  • A. Bespalov, D. Praetorius, L. R., and M. Ruggeri, Goal-Oriented error estimation and adaptivity for elliptic PDEs with parametric or uncertain inputs, Comput. Methods Appl. Mech. Engrg., vol. 345, pp. 951-982, 2019; download (Elsevier page).
  • A. Bespalov and L. R., Efficient adaptive algorithms for elliptic PDEs with random data, SIAM/ASA J. Uncertain. Quantif., vol.6, no. 1, pp. 243-272, 2018; download (SIAM page).
  • S. Cacace, E. Cristiani, and L. R., A level set based method for fixing overhangs in 3D printing, Appl. Math. Model., vol. 44, pp. 446-455, 2017; download (Elsevier page).
    The research has been mentioned by Le Scienze, the Italian edition of Scientific American; link (in Italian). Also, see the following presentation.

Projects

During my PhD, I have extensively contributed to two software packages for solving (deterministic and parametric) elliptic partial differential equations:

  • Triangular IFISS (T-IFISS) is a MATLAB toolbox for solving deterministic steady-state diffusion problems on general 2D polygonal domains using the Finite Element Method. It includes Galerkin approximations on triangular meshes, built-in a posteriori error estimation techniques, and adaptive algorithms with local mesh refinement.
    • Implementing built-in a posteriori error estimation of the computed solutions;
    • Implementing fast (vectorised) mesh-refinement routines and self-adaptive algorithms;
    • Improving runtime performances and efficiency.
  • Stochastic T-IFISS extends the core version of T-IFISS and includes a posteriori error estimation techniques and adaptive algorithms (including goal-oriented adaptive algorithms) for computing stochastic Galerkin approximations of steady-state diffusion equations with parametric uncertain coefficients.
    • Extending the core version of Triangular IFISS to cover stochastic approximations of PDEs with random data;
    • Implementing the associated adaptive algorithms, including goal-oriented adaptivity.

See LEBREF2D, a MATLAB package for the mesh refinement of 2D domains based on an efficient usage of MATLAB built-in functions and vectorization. It contains the core mesh-refinement routines in the adaptive algorithms of Stochastic T-IFISS.