“…Alternatively, there exist approaches to directly represent the PDE solutions via NNs. To mention a few: [9,10] minimize the strong form of the residual via collocation methods; [11] proposes a Deep Ritz Method (DRM) for symmetric and positive definite problems; [12] introduces a Deep Fourier Method; and [13][14][15][16][17] propose (Petrov-)Galerkin frameworks in the context of trial NNs with test functions belonging to linear spaces. All these NN-based methods exhibit multiple features but also present several limitations (see, e.g., [18][19][20][21]).…”