A verified open-access AI-based chemical microparticle image database for in-situ particle visualization and quantification in multi-phase flow

Abstract

Deep learning-based image segmentation and classification had been previously proposed for the in-situ measurement of microcrystal size and shape, which showed great potential to expand as a process analytical technology (PAT) to chemical multi-phase flow processes. In this study, an open-access chemical microparticle image database (open-CMD) was established containing particle agglomerations (A), bubbles (B), crystals (C), and droplets (D) in various chemical multi-phase flow scenarios. The advanced neural network, Mask R-CNN, coupled with 2,500 labeled images containing more than 50,000 labeled particles in open-CMD, was trained to build the ability of target particle segmentation and classification. The training results indicated that a data augmentation strategy could significantly improve the accuracy (<3.8 % average precision) of the trained models, which were named MicropNet+ and MicropNet according to whether the augmented data was used for training or not. Based on the superior capability of MicropNet+, multidimensional particle descriptors were extracted, and further, the degree of agglomeration and agglomeration distribution were defined and quantified. Then, two classical multi-phase flow processes, crystallization and emulsification, were analyzed using the MicropNet+ model, in which the agglomeration degree and distribution (Cin A) of succinic acid crystals and the relative number and diameter (Deq) of droplets were analyzed quantitatively under different operations conditions. It was concluded that the well-trained MicropNet+ model has high accuracy and efficiency in microparticle segmentation and classification. At last, the open-CMD database and the MicropNet+ model were released to inspire potential applications in chemical multi-phase flow areas.