Interlocking mechanism design based on deep-learning methods

Abstract

Biological structural systems such as plant seedcoats, beak of woodpeckers or ammonites shells are characterized by complex wavy and re-entrant interlocking features. This allows to mitigate large deformations and deflect or arrest cracks, providing remarkable mechanical performances, much higher than those of the constituent materials. Nature-inspired engineering interlocking joints has been recently proved to be an effective and novel design strategy. However, currently the design space of interlocking interfaces relies on relatively simple geometries, often built as a composition of symmetric circular or elliptical sutured lines. In the present contribution it is shown that deep-learning (DL) methods can be leveraged to enlarge the design space. Accurate and fast assessments of stiffness, strength and toughness of interlocking interfaces, generated through a cellular automaton-like method, can be obtained using a convolutional neural network trained on a limited number of finite element results. A simple application of a DL model for the recognition of interlocking mechanisms in 2-D interfaces is introduced. It is also shown that DL models, pre-trained on small resolution geometries, can accurately predict structural properties on larger design spaces with relatively small amounts of new training data. This work is addressed to give new insights into the study and design of a new generation of advanced and novel interlocked structures through data-driven methods.