Neural networks for automated soft-material mesophase discovery from simulation and experimental microscope data

Abstract

Three-dimensional network structures (Nets) formed by self-assembly of molecular amphiphiles, block oligomers, and block polymers are desirable for numerous applications, but Nets occupy only a relatively small window in the volume fraction-temperature phase diagram. Molecular simulations offer opportunities to screen block oligomer chemistries and architectures that promote Net formation. However, unambiguous identification of ordered network morphologies formed in molecular simulations is nontrivial. Previously, a PointNet model was used to detect globally ordered morphologies formed by linear diblock, linear triblock, and 3-arm and 4-arm star-block oligomers, and it also allows for the discovery of emerging ordered patterns from nonequilibrium systems. With the evolution and stabilization of the molecular dynamics simulations for Nets, including double gyroid (DG), double diamond (DD), single gyroid (SG), and double primitive (DP), and identification of new subclasses, we extend the previously developed Pointnet model for detection of nonlocal ordered morphologies of complex block oligomers. PointNet is a novel end-to-end trainable deep architecture used initially for 3D object detection, and it is one of the most popular point cloud classification methods in real applications. However, PointNet extracts global features but it ignores to predominantly capture fine local features, which limits its recognition ability of the fine-grained model of point clouds. How to build a local feature extractor to capture richer relevant interpoint information is the primary goal of this research. Recently, transformer techniques have successfully classified objects with image inputs, inspiring us to transplant transformers into our PointNet model. We exploit the self-attention mechanism to design a refined feature extractor and capture local geometric information. Thereby, rather than processing the whole point cloud, the network effectively learns where to look to find regions of interest. As a result, the number of points to be processed and inference time might reduce. The preliminary results show that the new network improves classification tasks and marginally outperforms our earlier version of PointNet developed by Shen et al. regarding per-class accuracy and stability.