A Hyper-Reduced Neural Network-Augmented Semi-smooth Newton Method for Nonlinear Parametric Variational Inequalities
We propose a model order reduction framework for nonlinear parametrized variational inequalities arising in computational mechanics. The high-dimensional model is written in mixed primal-dual form with projection-based complementarity conditions, leading to nonlinear nonsmooth algebraic systems solved by a semi-smooth Newton method in primal-dual form. On this basis, reduced models are constructed by proper orthogonal decomposition (POD) of both primal and dual solution snapshots, and the resulting reduced systems are solved by semi-smooth Newton iterations in the reduced space. To address cases where low-dimensional linear spaces provide limited approximation efficiency, we introduce a neural-network-augmented reduced model. Two feedforward networks learn corrections in the truncated POD coordinates of the primal and dual variables, defining a nonlinear manifold approximation that is embedded directly in the semi-smooth Newton iterations. The online cost associated with high-dimensional residual evaluations is reduced through hyper-reduction, using a sparse cubature approach based on greedy nonnegative least squares. Particular attention is paid to the interaction between hyper-reduction and the learned nonlinear manifold. The proposed methodology is assessed on two nonlinear variational inequalities with distinct sources of nonlinearity: a two-dimensional obstacle problem with a cubic nonlinearity in the state equation, and a three-dimensional frictional contact problem in which the Coulomb law induces a nonlinear projection in the constraint equation. Numerical results compare the high-dimensional model, the linear reduced model, the neural-network-augmented reduced model, and their hyper-reduced variants, demonstrating accurate approximations with substantial reductions in online computational cost.