This methodologically oriented project is devoted to the development and numerical analysis of learning algorithms for (typically large) data sets exploiting a priori knowledge in terms of an auxiliary physical model formulated as partial differential equation.