Machine learning approach to inverse design of topography transformations in liquid crystal elastomer coatings

While natural selection provides living species with complex design of soft systems, engineering design of actuating soft materials to achieve target shape-morphing is a theoretical/computational challenge [1-2] . Our research aim is to address this challenging inverse problem to predict the behavior of liquid crystal elastomer (LCE) coatings using machine learning (ML). We present a tabular supervised machine learning approach to design LCE thin coatings that transform on stimulus to produce a desired topography. We generate a training dataset by running a nonlinear finite element elastodynamics "forward" simulation for 1500 LCE samples with different microstructures. We use 80% of the dataset to train a stacked ensemble regression model using the AutoGluon [3] framework to apply multiple modeling methodologies, including tree-based and deep learning algorithms, to maximize the performance. We evaluate the resulting model for the remaining 20% of the dataset. Finally, we demonstrate the capability of our approach to predict the parameters needed to create switchable surfaces of LCE coatings that transform from a flat profile to produce an array of suction cups that resemble octopus suckers. [1] F. D. Moura Neto and A. J. Da Silva Neto, "An introduction to inverse problems with applications" Springer 2013 [2] H. Aharoni et al, PNAS 2018 [3] auto.gluon.ai