(178c) Deep Learning-Based Sensor for Crystal Size and Contour Characterization

In the field of crystallization, practitioners do not have access to sensors that can make real-time, accurate measurements of crystal size in high-density slurries. Several paths have been explored to solve this problem but most of them have come up short when applied to industrial settings. One path that has shown some promise is video microscopy and image analysis. In this approach, a probe is inserted into the crystal slurry and a microscope captures images. The images are sent to an imaging software, where the individual crystals are segmented using variations of a two-step algorithm: morphological treatment followed by particle segmentation. This approach works well when the particles are adequately separated but breaks down in dense slurries. Particles in dense slurries have their shapes distorted by attrition and experience significant overlap due to surface forces. This obscures the intra-particle boundaries and complicates the task of segmenting crystal particles. Detecting particles in such an environment requires a sensor that is robust to large variations in particle size, shape and visibility.

In recent years, the field of computer vision has switched over from traditional feature extraction-based algorithms to more powerful tools like deep learning (DL) and artificial intelligence (AI). This transition has led to impressive breakthroughs in a variety of computer vision challenges like image labelling, object detection and instance segmentation. The success of these tools has spurred researchers to revisit unresolved industrial problems and discover new solutions.

In this work, we propose a deep learning-based sensor that solves the problem of crystal detection in high-density slurries using the Mask RCNN, a state-of-the-art CV model. The sensor functions in three stages. In the first stage, it analyzes the image pixels and extracts meaningful features like edges, textures, patterns, parts and objects. In the second stage, it uses this information to weed out the background and identify locations in the image that enclose crystal particles. In the final stage, the coordinates of the locations are iteratively refined to tightly enclose the particles and segmentation masks are generated to isolate the contours of individual particles. This information is then used to precisely quantify the shape characteristics of the crystals and generate the size distribution.

The Mask RCNN architecture used in the sensor is powered by a ResNet-101 backbone and incorporates approximately 63 million parameters. All the parameters were pre-trained on the COCO dataset and one-third of them were further fine-tuned on the crystal dataset. Our dataset has 100 labeled training images and 20 labeled validation images, each with at least 70 crystal particles. The original images were resized to 1024x1024 pixels and extensive data augmentation was used to prevent overfitting. The model converges in approximately 30 epochs, takes 4 hours to train and can process images up to a speed of 5 FPS on a GPU.  It shows promising results in segmenting crystal particles in a wide variety of conditions. Figure 1 shows the results for the anti-solvent crystallization of sodium chloride in water using ethanol as anti-solvent in an experimental bench-scale semi-batch crystallizer.

Figure 1. (a) Crystal segmentation results for each crystal instance. (b) The crystal size distribution drawn from the segmentation results.