Accelerated design of architectured ceramics with tunable thermal resistance via a hybrid machine learning and finite element approach

Abstract

Topologically interlocked architectures can transform brittle ceramics into tougher materials, while making the material design procedure a cumbersome task since modeling the whole architectural design space is not efficient and, to a degree, is not viable. We propose an approach to design architectured ceramics using machine learning (ML), trained by finite element analysis data and together with a self-learning algorithm, to discover high-performance architectured ceramics in thermomechanical environments. First, topologically interlocked panels are parametrically generated. Then, a limited number of designed architectured ceramics subjected to a thermal load is studied. Finally, the multilinear perceptron is employed to train the ML model in order to predict the thermomechanical performance of architectured panels with varied interlocking angles and number of blocks. The developed feed-forward artificial neural network framework can boost the architectured ceramic design efficiency and open up new avenues for controllability of the functionality for various high-temperature applications. This study demonstrates that the architectured ceramic panels with the ML-assisted engineered patterns show improvement up to 30% in frictional energy dissipation and 7% in the sliding distance of the tiles and 80% reduction in the strain energy, leading to a higher safety factor and the structural failure delay compared to the plain ceramics.