From structured data to evolution linear partial differential equations (Q2222252)

In this paper, a set of direct methods is proposed for constructing linear evolution PDEs from data with experimental values of the solution of the unknown equation, and along with the corresponding positions and times. It is assumed that the searched equation is a linear evolution equation with non-constant coefficients. This type of data is typically available ``in-lab'' experiments, that makes the proposed approach potentially relevant from a practical point of view. Remarkably, machine learning paradigms are not directly used. Using pseudospectral methods and symbolic representation of pseudo differential operators, an inverse problem is solved by assuming that the data is structured and randomly weakly perturbed. Some promising results are obtained for sufficiently small amplitude perturbations. Extension to quasilinear equations is discussed. The proposed methodology is validated by several numerical experiments for a one-dimensional linear equation with space dependent coefficients, a first order quasilinear equation, a fractional equation in 2D, and a linear equation with non-constant (space-time dependent) coefficients.