Quantifying Nanoparticle Assembly States in Polymer Matrix Through Deep Learning

There is great interest in controlling the spatial dispersion of inorganic nanoparticles (NPs) in an organic polymer matrix, because this centrally underpins the property enhancements obtained from these hybrid materials. Currently, qualitative information on NP spatial distribution is obtained by visual inspection of transmission electron microscopy (TEM) images. Quantitative information is only indirectly obtained through the use of scattering probes such as small angle X-ray/neutron scattering. While the main challenge, that scattering probes operate in reciprocal space, can be remedied by Fourier inverting the data into real space, a much harder issue is deconvolving the contribution of the particle form factor (which is affected by the details of the NP size and shape) from the structure factor which contains information on the NP spatial distribution. These problems become acute when we deal with the popular topic of NPs grafted with polymer chains, because the polymeric corona, and hence the particle form factor, becomes context dependent and hard to quantify. To make progress, we develop and apply a deep-learning based image analysis method to quantify the distribution of spherical NPs in a polymer matrix directly from their real-space TEM images. A dataset of NP detection (DOPAD) is built by manually labeling particle positions on experimental TEM images of diverse polymer composite systems. A convolutional neural network (CNN) object detection model is then trained on DOPAD. Together with sliding-window and merging algorithms, an automated pipeline is established, which takes a large TEM image as input and extracts NP locations and sizes. We validate the structural information resulting from this method against SAXS derived structural information for NPs ordered by polymer crystallization, and then use it to distinguish between different states of the assembly of polymer grafted NPs in a polymer matrix achieved by using their surfactancy. We show that this data-rich protocol allows us to draw critical facets of experimental behavior which have previously not been accessible. The DOPAD dataset, Python source code and trained model are shared on GitHub.