Deep learning-based non-intrusive detection of instabilities in formulated liquids

Abstract

Stability is a key property of formulated liquids for industrial applications. Before their commercialization, a solid assessment of stability under various operative conditions must be carried out. Traditionally, this is performed by expert researchers that observe the liquid over time and point out the possible occurrence of instabilities. However, this is a costly and time-consuming process. Here, we investigate the potential of deep learning approaches for automatic image-based assessment of formulated liquid stability. Leveraging a recently developed dataset, comprising thousands of images of formulated liquids stored in transparent jars, we implement and test several state-of-the-art Convolutional Neural Networks (CNNs) with different loss functions and augmentation strategies. Experiments prove the effectiveness of this non-invasive approach opening the way to further applications.