Robust topology optimization with low rank approximation using artificial neural networks (Q2667321)

The aim of the paper is a low-rank approximation approach for topology optimization (under uncertainty) of parametrized linear elastic structures. The elastic response of the structure without uncertainty is characterized by the following boundary value problem: \begin{align*} \nabla\cdot \sigma (x)+ \textbf{b}(x)&=0 & &\forall x \in D,\\ \sigma(x)\textbf{n}&=n(x)& &\forall x \in \Gamma_N,\\ u(x)&=0 & &\forall x \in \Gamma_D, \end{align*} which can be presented in the form of a linear elliptic partial differential equation \begin{align*} -\nabla\cdot (C(x,\rho)\nabla u(x))&=f(x) & &\forall x \in D,\\ u(x)&=0& &\forall x \in \partial D. \end{align*} With uncertainty this translates to finding find a stochastic function \(u=u(y,x)\), where \(y\in \mathcal{Y}\subset (\Omega, \mathcal{A},\mathcal{P})\) is the probabilistic variable, such that for \(\mathcal{P}\)-almost everywhere \(y\in \mathcal{Y}\) the following equation holds: \begin{align*} \nabla\cdot (C(y,x)\nabla u(y,x))&=f(x) & &\forall (y,x) \in \mathcal{Y}\times D,\\ u(y,x)&=0& &\forall (y,x) \in \mathcal{Y}\times\partial D. \end{align*} The usual practice to reduce the enormous amount of computation time is to develope reduced order models for an efficient yet accurate approximation. The key idea is to identify (learn) the parametric variations from an inexpensive model and use this knowledge to reconstruct parametric high-fidelity samples. This is achieved by adopting an artificial neural network surrogate as a powerful function approximator and surrogate. As numerical examples the MBB beam, a mechanical inverter and a three-dimensional topology optimization, i.e., the task of designing a robust heat sink which minimizes the thermal compliance, as a more computationally demanding problem are presented.