Deep Learning Enhanced Snapshot Generation for Efficient Hyper-reduction in Nonlinear Structural Dynamics

This study presents a novel approach to enhancing hyper-reduction in nonlinear structural dynamics by utilizing the predictive capabilities of stacked Long Short-Term Memory (LSTM) neural networks. Hyper-reduction methods are crucial for overcoming the limitations of traditional model order reduction techniques, particularly in accurately capturing the nonlinear behavior of internal force vectors in complex structures.

The proposed technique employs stacked LSTM neural networks to generate training snapshots for the Energy Conserving Mesh Sampling and Weighting (ECSW) hyper-reduction method. By training the model on a well-defined dataset, we achieve an impressive accuracy of 97.5%. The effectiveness of our method is demonstrated through a geometrically nonlinear dynamic analysis of a leaf spring, resulting in only a 3.24% error when compared to full simulation results. This study emphasizes the potential of deep learning techniques in improving hyper-reduction methods and underscores the importance of computational efficiency in simulations of complex structural dynamics.

The findings reveal significant advancements in the application of deep learning for hyper-reduction methods, showcasing the ability to accurately model nonlinear structural behaviors while maintaining computational efficiency. This research contributes valuable insights into the integration of advanced machine learning techniques within the field of structural dynamics.