Stability assessment of liquid formulations: A deep learning approach

[Display omitted] •A non-intrusive method for stability assessment of formulated liquids is proposed.•A dataset of images of liquid sample vials is acquired and labeled by experts.•A deep learning-based classifier is trained and tested on the dataset.•The classifier reliably detects the onset of instabilities in liquid samples.•Further analyses show that it can also locate the instability in the vial. Formulated liquids, such as detergents and fabric softeners, are of great relevance in the industry of everyday products. Although thermodynamically unstable, these liquids must guarantee a minimum stable ‘shelf life’ for commercialization. Therefore, the industry devotes a major effort to assess their stability, and to develop reliable instability detectors. Among the many approaches proposed to this end, solutions based on the visual inspection of samples present numerous advantages: they are cheap, easy to perform, non-intrusive, repeatable. However, conventional image-processing-based detectors do not ensure a sufficient reliability. In this paper, we propose to use deep learning-based classifiers, thus we explore several architectural choices to guarantee the desired performance with a limited computational cost, despite the relatively small datasets available for this task. Experimental results are extremely encouraging and show that deep learning is a reliable and very promising solution for detecting instabilities in formulated liquids.