A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized PDEs (Q831238)

Non-intrusive deep learning (DL) algorithms are proposed to build low-dimensional reduced-order models (ROMs) to overcome computational bottlenecks shown by classical, linear projection-based ROM techniques when dealing with problems featuring wave-type coherent structures propagating over time. DL-ROM technique approximates both the solution manifold of a given parametrized time-dependent nonlinear PDE by means of a low-dimensional nonlinear trial manifold, and the nonlinear dynamics of the generalized coordinates on the reduced trial manifold. The nonlinear trial manifold and the reduced dynamics are learnt in a non-intrusive way by means of neural networks. The numerical results obtained for Burger's equation, linear transport equation, and monodomain equation show that DL-ROMs provide approximations to the parameterized PDEs that are orders of magnitude more accurate than the ones provided by linear ROMs, for the same reduced dimension. The proposed DL-ROM is computationally more efficient than the linear ROMs with global proper orthogonal decomposition (POD) bases and nonlinear ROM local POD bases, when applied to the linear transport equation or nonlinear diffusion-reaction PDEs coupled to ODEs.