A mesh-free method using piecewise deep neural network for elliptic interface problems (Q2141617)

The aim in this paper is to investigate the use of deep learning methods to solve interface problems. One focuses on the second-order elliptic interface problem and the research follows the following scheme: the interface problem is reformulated by means of a least-square (LS) approach as a minimization problem and two deep neural network structures (DNN) are built to approximate the solution on sub-domains. The following second-order scalar elliptic interface problem is considered: \[ -\nabla.(\beta(\mathcal{X})\nabla u)=f,\text{ in }\Omega_1\cup\Omega_2, \] \[ [[u]]=g_i\text{ on }\Gamma, \] \[ [[\beta(\mathcal{X})\nabla u.\mathcal N]]=g_f\text{ on }\Gamma, \] \[ u=g_D,\text{ on }\partial\Omega \] where the interface \(\Gamma\) is closed and divides domain \(\Omega\) in two disjoint sub-domains \(\Omega _1\) (the interior) and \(\Omega _2\) (the exterior). One defines the (LS) functional incorporating the interface as well as boundary conditions \(\mathcal {J}\)(v;\(g_j\),\(g_f\),\(g_D\),f) and one looks for its minimum for v\(\in\) \(H^1\)(\(\Omega\)). The construction of the two neural networks to approximate the solution on both sub-domains \(\Omega_1\) and \(\Omega _2\) is discussed in the third section of the article. One develops an adaptive deep LS algorithm. Results on numerical experiments as for instance on Sunflower Shape Interface, Sphere Shape Interface, Heart Shape Interface, Circle Interface with High Contrast Coefficients, Flower Shape Interface are presented in the fourth section of the paper.