FastSVD-ML–ROM: A reduced-order modeling framework based on machine learning for real-time applications

Digital twins have emerged as a key technology for optimizing the performance of engineering products and systems. High-fidelity numerical simulations constitute backbone design, providing an accurate insight into complex However, large-scale, dynamic, non-linear models require significant computational resources are prohibitive real-time digital twin applications. To this end, reduced order (ROMs) employed, to approximate high-fidelity solutions while accurately capturing dominant aspects physical behavior. The present work proposes new machine learning (ML) platform development ROMs, handle large-scale problems dealing with transient nonlinear partial differential equations. Our framework, mentioned $\textit{FastSVD-ML-ROM}$, utilizes $\textit{(i)}$ singular value decomposition (SVD) update methodology, compute linear subspace multi-fidelity during simulation process, $\textit{(ii)}$ convolutional autoencoders dimensionality reduction, $\textit{(iii)}$ feed-forward neural networks map input parameters latent spaces, $\textit{(iv)}$ long short-term memory predict forecast dynamics parametric solutions. efficiency $\textit{FastSVD-ML-ROM}$ framework is demonstrated 2D convection-diffusion equation, problem fluid around cylinder, 3D blood flow inside arterial segment. accuracy reconstructed results demonstrates robustness assesses proposed approach.