Designing architectured ceramics for transient thermal applications using finite element and deep learning

Abstract Topologically interlocking architectures have demonstrated the potential to create durable ceramics
with desirable thermo-mechanical properties. However, designing such materials poses
challenges due to the intricate design space, rendering traditional modeling approaches ineffective
and impractical. This paper presents a novel approach to designing high-performance architectured
ceramics by integrating machine learning (ML) techniques and finite element analysis (FEA)
data. The design space of interlocked architectured ceramics encompasses tiles with varying angles
and sizes. The study considers three configurations 3 × 3, 5 × 5, and 7 × 7 arrays of tiles
with five sets of interlocking angles (5◦, 10◦, 15◦, 20◦, and 25◦). By training ML models, specifically
convolutional neural networks (CNNs) and multilayer perceptrons (MLPs) using FEA simulation
data, we establish correlations between architectural parameters and thermo-mechanical characteristics.
A grid comprising all possible designs was generated to predict high-performance
architectured ceramics. This grid was then fed into the networks that were trained using results
from the FEA simulation. The predicted results for all possible interpolated designs are utilized
to determine the optimal structure among the configurations. The goal is to identify the optimal
interlocked ceramics that minimize the out-of-plane deformation for thermal shielding and maximize
heat absorption for heat sink applications. To validate the performance of the outcomes,
FEA simulations were conducted on the best predictions obtained from both the MLP and CNN
algorithms. Despite the limited amount of available simulation data, our networks demonstrate
effectiveness in predicting the transient thermo-mechanical responses of potential panel designs.
Notably, the optimal design predicted by CNN led to ≈30% improvement in edge temperature.