(46f) Real-Time Particle Size Estimation for Crystallization Processes Through GPU-Based Multivariate Image Analysis

Real-time Particle Size Estimation for Crystallization Processes through GPU-based Multivariate Image Analysis

Lau Mai Chan, Rajagopalan Srinivasan

Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore, 10 Kent Ridge Crescent, Singapore 119260, Singapore

Abstract

The capability to estimate crystal size distribution in real-time is important for effective control and optimization of particulate processes. Appropriately controlled crystal size distribution not only ensures high efficiency of downstream operations like filtering, drying and formulation, it also safeguards the efficacy of final product which is in crystal form. In order to achieve fast online measurement of crystal size, automated image analysis has recently been developed (Sarkar 2009). The methods have been shown to be quite accurate in terms of the particle size distribution estimated in real-time.  However, this comes at the cost of computational efficiency. The major objective of the current work is to improve the computational efficiency of multivariate image analysis (MIA) such that the computational speed is at least on par with the image generation speed.

In this work, a Particle Vision and Measurement (PVM) probe which captures in situ images at the speed of 2 images / second provides the benchmark for expected performance. Multivariate image analysis consists of various tasks including extraction of image features, construction of statistical image model, segmentation based on blank background image, post-segmentation image analysis and boundary refinement. Our profiling of the computational costs revealed that the first two tasks, feature extraction and construction of statistical image model (which essentially involves scaling and principal component analysis), account for 90% of the total computation time. We improve the computational efficiency of MIA by implementing the aforementioned computationally expensive tasks in a Graphics Processing Unit (GPU). General purpose computing on graphics processing unit (GPGPU) allows computational efficiency of non-graphical applications to be improved. This is realized by effectively mapping computational operations to massive processing cores of GPU on which they can be executed concurrently. In MIA, the two computational tasks, feature extraction and scaling, are inherently rich in data parallelism i.e. same operations performed on different data (or pixels, in this case); therefore they are well suited to parallel computing on a GPU.

The proposed GPU-MIA algorithm has been implemented and tested on a case study, which involves batch crystallization of monosodium glutamate monohydrate (MSG) for a period of 23 hours. Computational studies show that the average time for extracting particle size information from images has been reduced from the original 0.70 seconds per image to 0.11 seconds per image – an overall improvement of 6.1 times. Therefore the proposed GPU-MIA implementation has achieved the desired target of 0.5 seconds per image. The details of the GPU algorithm and the computational studies will be presented in this paper.

Reference

Sarkar, D., Doan, X.-T., Ying, Z., Srinivasan, R. "In situ particle size estimation for crystallization processes by multivariate image analysis." Chemical Engineering Science, 64 (1), 2009: 9-19.