Selective boundary condition reduction via learned error gating (Q7288239)

This paper presents a framework for learning when to use simplified boundary conditions in parametric PDEs, reducing computational cost while maintaining accuracy: neural networks estimate domain and boundary errors of paired solutions, replacing the more expensive condition with the simpler one if both predicted errors meet prescribed tolerances. We focus on singular limits and evaluate the method on a galvanic corrosion problem and other nonlinear stationary and evolution problems.