Semi-auxetic sandwiches are laminates composed of conventional and auxetic laminae. They demonstrate properties that are absent in their constituents, e.g., the in-plane effective elastic modulus of the sandwich structure surpasses the rule of mixture while the out-of-plane value cannot reach the inverse rule of mixture2,3. Traditional design of such metamaterials (engineered materials with extraordinary properties) requires experience and computational algorithms that often use too expensive finite element simulations to estimate the mechanical response.
Deep learning techniques can accelerate the forward design process by using artificial neural networks (ANNs) for geometry generation and obtaining effective properties4. More importantly, a trained ANN can be used to inverse design metamaterials1, see the Figure. Particularly, autoencoders encode the high-dimensional input into a lower dimensional latent space, which represents the likelihood of the input data. Thus, it can be decoded to generate various types of synthetic data5 and new microstructures when generative models are used.
This project aims to use autoencoders in inverse design of semi-auxetics for optimised energy absorption. The objectives are (1) to generate the training dataset using the finite element method; (2) to propose an ANN model that includes re-entrant honeycomb as the input geometry and elastic modulus, Poisson’s ratio, and toughness as outputs; and (3) to inverse-design a geometry with maximum toughness.
References
[1] X. Zheng, X. Zhang, T.-T. Chen and I. Watanabe: Advanced materials (Deerfield Beach, Fla.), 2023, 35, (45), e2302530.
[2] T.-C. Lim and U. Rajendra Acharya: Physica Status Solidi (b), 2011, 248, (1), 60–65.
[3] T.-C. Lim: European Journal of Mechanics - A/Solids, 2009, 28, (4), 752–756.
[4] M. Mohammadnejad, A. Montazeri, E. Bahmanpour and M. Mahnama: Thin-Walled Structures, 2024, 200, 111927.
[5] J. Mounayer, S. Rodriguez, C. Ghnatios, C. Farhat and F. Chinesta: ‘Rank Reduction Autoencoders’ 22/05/2024.
