Machine learning for tailored multifunctional properties of lightweight metamaterials

Abstract

Metamaterials are man-made materials with periodic cellular architectures, offering properties that are defined by geometry rather than chemical composition. This allows significant potential for lightweighting and tunability of mechanical responses that can be tailored to specific practical loading scenarios. Furthermore, additional levels of multifunctionality can be activated and tuned according to requirements such as sound and energy absorption. The principal aim of this thesis is to explore the effectiveness of machine learning for designing metamaterial architectures that maximise mechanical performance while performing multi-objective optimisations to balance trade-offs between competing design objectives, enabling enhanced multifunctionality. A comprehensive literature review has been carried out to explore the current state of the art in architected metamaterials. Machine learning offers a vast array of potential applications across engineering disciplines. A selection of five studies was presented in which workflows were developed to allow tailoring of material properties by artificial intelligence. An inverse convolutional neural network was used to fine-tune sound absorption responses of micro-perforated panel sound absorber metamaterials for maximum absorption amplitude and range. The process was then applied to tune mechanical responses of truss-based lattice structures for enhanced structural efficiency. Networks were then trained to perform inverse tailoring of combined sound absorption and structural performance of a quarter-wavelength resonator labyrinth unit cell. An alternative approach was explored that uses neural networks to predict performance objectives in combination with evolutionary optimisation algorithms. This method was used to synthesise material architectures for enhanced stiffness, strength and high velocity impact resistance. Both inverse and predictor approaches were used to tailor layup sequences for optimal stress responses in composites. An early proof of concept was also obtained for use of artificial intelligence as a faster alternative to finite element analysis in parametric stress analysis across engineering design practices. The developed network architectures and methodologies may be interchangeable with other problems of similar complexity and can theoretically be applied to any multi-variable design problem for which a dataset can be produced to train a machine learning model.

More information

Identifiers

ORCID for Peter Anthony Hawes: orcid.org/0009-0006-0909-2097

ORCID for Michele Meo: orcid.org/0000-0003-1633-8930

ORCID for Jie Yuan: orcid.org/0000-0002-2411-8789

Catalogue record

Export record

Altmetrics