Physics-informed machine learning in asymptotic homogenization of elliptic equations

Abstract

We apply physics-informed neural networks (PINNs) to first-order two-scale periodic asymptotic homogenization of the property tensor in a generic elliptic equation. The problem of lack of differentiability of property tensors at the sharp phase interfaces is circumvented by making use of diffuse interface formulation. Periodic boundary conditions are incorporated strictly, through the introduction of an input-transfer layer (Fourier feature mapping), in which the sine and cosine of the inner product of position vectors and reciprocal lattice vectors are considered. This, together with the absence of Dirichlet boundary conditions, results in a lossless boundary condition application. The only loss terms are then due to the differential equation itself, which removes the necessity of scaling the loss entries. In demonstrating the formulation's versatility based on the reciprocal lattice vectors, crystalline arrangements defined with Bravais lattices are used. We also show that considering integer multiples of the reciprocal basis in the Fourier mapping leads to improved convergence of high-frequency functions. We consider applications in one, two, and three dimensions. Periodic composites, composed of embeddings of monodisperse inclusions in the form of disks/spheres in the two-/three-dimensional matrix, are considered. For demonstration purposes, stochastic monodisperse disk arrangements are also considered.