Engineering Mechanics

Proceedings Vol. 31 (2025)

ENGINEERING MECHANICS 2025

Copyright © 2025 Institute of Theoretical and Applied Mechanics of the Czech Academy of Sciences, Prague

 

Raisinger J., Novák L., Eliáš J.
pages 169 - 172, full text

Multiscale homogenization enables the calculation of the macroscopic stress response of a discrete periodic representative volume element (RVE) of a (quasi) brittle materials subjected to a macroscopic strain, accurately capturing their nonlinear inelastic behaviour. Wider application of the RVE as a microstructural constitutive model is hindered by its high computational cost. This paper investigates two data-driven approaches to approximate the effective response of the RVE. 3D RVE is loaded by uniaxial macroscopic strain and the response is approximated using polynomial chaos expansion (PCE) and neural networks. Both approaches incorporate a predefined history variable to incorporate the path dependency, while the second also tries to leverage recurrent neural networks (RNNs) to learn the load path dependency autonomously. The results show that the RNN can reach the same accuracy as a feed-forward network with a history variable given enough training data. The PCE provides good results but does not reach the precision of the other methods.

Ownership of copyright in original research articles remains with the Authors, and provided that, when reproducing parts of the contribution, the Authors acknowledge and/or reference the Proceedings, the Authors do not need to seek permission for re-use of their material.

All papers were reviewed by members of the scientific committee.