Speaker
Description
Damage simulations play a critical role in assessing structural integrity and understanding material failure under various conditions. However, their computational cost is often a significant challenge, especially in scenarios requiring repeated simulations such as design optimization, uncertainty quantification, and real-time decision-making. Traditional model reduction techniques such as the proper orthogonal decomposition and especially hyper-reduction methods [1] can reduce the simulation time significantly but often necessitate intrusive modifications to existing simulation codes and are bound to the classical finite element framework. In this contribution, an autoencoder-based non-intrusive model reduction framework for gradient-extended damage simulations [2] is introduced. Autoencoders, a class of neural networks designed for dimensionality reduction and feature extraction, are employed to construct a compact latent space representation of high-fidelity simulation data. The autoencoder can be trained to capture complex, nonlinear relationships within the data, enabling significant reductions in computation time while maintaining high accuracy [3]. It is also easily possible to incorporate real-world data directly into the reduced order model because the data does not have to be transformed to e.g., Neumann or Dirichlet boundary conditions. The proposed methodology is validated using numerical examples involving complex structural damage simulations. Different neural network structures and techniques are investigated with respect to their influence on accuracy. Results demonstrate the framework's ability to achieve orders-of-magnitude reduction in computational time while maintaining high accuracy in predicting damage evolution and structural behavior. The model's robustness and efficiency make it a promising tool for applications requiring rapid simulation capabilities, such as real-time monitoring, predictive forecasting, and uncertainty quantification in the context of digital twins.
[1] Farhat, C., Avery, P., Chapman, T., & Cortial, J. (2014). Dimensional reduction of nonlinear finite element dynamic models with finite rotations and energy‐based mesh sampling and weighting for computational efficiency. International Journal for Numerical Methods in Engineering, 98(9), 625-662.
[2] Brepols, T., Wulfinghoff, S., & Reese, S. (2020). A gradient-extended two-surface damage-plasticity model for large deformations. International Journal of Plasticity, 129, 102635.
[3] Simpson, T., Dervilis, N., & Chatzi, E. (2021). Machine learning approach to model order reduction of nonlinear systems via autoencoder and LSTM networks. Journal of Engineering Mechanics, 147(10), 0402106
