2522 Microsc. Microanal. 27 (Suppl 1), 2021 doi:10.1017/S1431927621008990 © Microscopy Society of America 2021 Learning to Estimate the Composition of a Mixture with Synthetic Data Cuong Ly 1 , Cody Nizinski 2 , Clement Vachet 3 , Luther McDonald 2 and Tolga Tasdizen 4 1 University of Utah - Scientific Computing and Imaging Institute, Salt Lake City, Utah, United States, 2 University of Utah, United States, 3 Scientific Computing and Imaging Institute, United States, 4 University of Utah - Scientific Computing and Imaging Institute, United States Identifying the precise composition of a mixed material is important in various applications. For instance, in nuclear forensics analysis, knowing the process history of unknown or illicitly trafficked nuclear materials when they are discovered is desirable to prevent future losses or theft of material from the processing facilities. Motivated by this open problem, we describe a novel machine learning approach to determine the composition of a mixture from SEM images. In machine learning, the training data distribution should reflect the distribution of the data the model is expected to make predictions for, which can pose a hurdle. However, a key advantage of our proposed framework is that it requires reference images of pure material samples only. Removing the need for reference samples of various mixed material compositions reduces the time and monetary cost associated with reference sample preparation and imaging. Moreover, our proposed framework can determine the composition of a mixture composed of chemically similar materials, whereas other elemental analysis tools such as powder X-ray diffraction (p- XRD) have trouble doing so. For example, p-XRD is unable to discern mixtures composed of triuranium octoxide (U3O8) synthesized from different synthetic routes such as uranyl peroxide (UO4) and ammonium diuranate (ADU) [1]. In contrast, our proposed framework can easily determine the composition of uranium oxides mixture synthesized from different synthetic routes, as we illustrate in the experiments. Fig. 1 shows an overview of our proposed approach, which is divided into three steps as highlighted with the dashed rectangles. For the first step (Fig. 1a), we develop a model to generate a synthetic dataset of mixed materials using real images of pure materials only. Our image synthesis model is based on the texture synthesis work proposed by Gatesy et al. [2], which generates a new image with the same texture as a given reference image by minimizing the differences between second-order statistics of features (Gram matrix) of the generated and reference images [2]. We define each pure material as a texture. A new synthetic mixture image is initialized as an image of white noise. We then minimize the differences between the Gram matrix of this image and the weighted sum of Gram matrices of the pure reference images. In the second step (Fig. 1b), we train a convolutional neural network (CNN) to estimate mixture compositions from SEM images using synthetic images generated from the first phase. After the training process, the mixture composition estimation model is deployed to estimate the composition of a mixture. This step is shown in Fig. 1c. We utilized mixtures of U3O8 made from ADU, uranyl hydroxide (UH), and sodium diuranate (SDU) to validate the proposed method. We refer to these U3O8 materials simply by their precipitation route, e.g., a U3O8 sample made from ADU is simply referred to as ADU for the rest of this paper. First, we used the image synthesis model to obtain the necessary dataset for training the mixture composition estimation model. In this experiment, the image synthesis model takes SEM images of ADU, UH, and SDU as input and produces images of the mixtures of prescribed percentages. Next, we used the synthetic images to train the mixture composition estimation model, which is tasked with estimating the percentage of ADU, UH, and SDU presence in an SEM image. Then, the mixture composition estimation model was tested on a set of real mixture images to validate the performance of the proposed framework. The set of real mixture images used for testing consists of six compositions: 100%ADU,100%UH, 100%SDU, 50%ADU- https://www.cambridge.org/core/terms. https://doi.org/10.1017/S1431927621008990 Downloaded from https://www.cambridge.org/core. IP address: 3.80.168.101, on 05 Nov 2021 at 12:00:22, subject to the Cambridge Core terms of use, available at