A machine learning approach to inverse design of programmable liquid crystal elastomers: morphing surfaces

Design of liquid crystal elastomer (LCE) devices requires solving an inverse problem: find the nematic director field that morphs a sample to a desired target shape when heated. We present a novel machine learning approach to address the inverse problem for LCE coatings on a rigid substrate, morphing to a target topography[1]. We solve 1500 forward problems via finite element methods [2] for various director configurations to form a training dataset. Next, we train a stacked ensemble regression model using the Autogluon framework [3]. In this autoML solution, multiple modeling methodologies are tried, and an ensemble model is constructed to maximize performance. Hyperparameter tuning is automatically handled by the API. Here 80% of the dataset was used to train the models including tree-based and deep learning algorithms. The prediction of the two parameters defining the director field on the remaining test dataset was evaluated. The ensemble model outperformed any individual model and could also predict configurations that departed from the initial geometry by adding noise. We discuss plans to extend this approach to a broader class of LCE device geometries. [1] Babakhanova 2020 doi:10.1063/5.0022193 [2] Gelebart 2017 doi:10.1038/nature22987 [3] auto.gluon.ai