Reduced Order Modelling (ROM), such as the Proper Generalized Decomposition (PGD) [1], offers accelerated solutions for partial differential equations (PDEs). However, their effectiveness often relies on parametrization, limiting applicability to fixed geometries geometries parametrizable by a couple of control points. To address this limitation, we propose integrating deep learning with ROM. Our method leverages the spatiotemporal separated variable representation of a PDE solution. We employ spatial modes from a PGD decomposition that offers a superior solution representation compared to eigenmodes of the structure . A first guess of the spatial component (spatial modes) is obtained using Graph Neural Networks (GNN) , allowing us to handle diverse and non-parametric geometries. The temporal component (temporal modes) is derived by projecting the PDE onto spatial modes from GNN. This ensures solving the PDE at a reduced cost while maintaining physical fidelity and allows the addition of new corrections using a greedy algorithm if it appears to be necessary. As a representative example, we focus on the dimensioning of aircraft seats in fast dynamics. Using deep learning on a database, we reuse accurate calculations from various engineering projects. We tackle non-parametric challenges in variable geometries, taking into account data sparsity. The uncertainty of deep learning-based solutions impacts certification and industry acceptance. By combining deep learning with industrial solvers, we can respond to the industry’s hesitations in adopting AI-based approaches while simultaneously offering the possibility of tackling non-parametric problems. REFERENCES [1] A. Daby-Seesaram, A. Fau, P.E- Charbonnel, D. N ́eron, A hybrid frequency-temporalreduced-ordermethod for nonlinear dynamics.Nonlinear Dynamics111(15), 2023 [2] J. Brandstetter, D. Worrall, M. Welling, Message Passing Neural PDE Solvers.arXiv:2202.03376, 2022