Hard-Constraint Projection

A method that adds a non-learnable projection layer to the final layer of a neural network, strictly projecting outputs onto the hyperplane (or constraint manifold) containing the exact solutions of discrete equations, so that physical constraints are satisfied exactly.