Generative Deep Learning-Aided Design of Flexible Molecular Crystals
Résumé
Organic molecular crystals capable of mechanical adaptation are poised to revolutionize soft advanced materials with potentially immense implications from optics to electronics and biomedicine. While these prospects have guided studies into both fundamental and performance aspects, currently available methodologies for the design of organic crystalline matter with specific mechanical properties based on classical crystal engineering principles lack reliability and generality in application, and this significantly limits consideration of organic crystals as materials of choice. To address this challenge, we apply deep learning models for the design of mechanically compliant organic crystalline materials. We introduce CrystalGAN, a deep generative framework based on a generative adversarial network (GAN), designed to efficiently generate flexible molecular crystals with desired mechanical properties. CrystalGAN leverages a graph convolutional network (GCN) to construct both the generator and discriminator of a Wasserstein GAN (WGAN), enhancing the validity of the generated molecules. A convolutional neural network (CNN) was trained to predict and discriminate the mechanical properties of unknown molecules, based on the data collected from the extant literature and compared with multilayer perceptron (MLP) with backpropagation algorithm. The CNN showed favorable performance with high accuracy in various computational evaluations and successfully predicted the mechanical response of the flexible crystals. The inferences are complemented by case studies that employed CrystalGAN and CNN to generate molecules that are expected to crystallize as flexible crystals with improved tableting properties. This work overcomes one of the current major challenges with the lack of discovery and prediction of organic crystals with specific mechanical properties.