Investigating the metastability of amorphous calcium carbonate by droplet microfluidics experiments using machine learning.

Amorphous calcium carbonate (ACC) plays an important role in the crystallization pathways of calcite and its polymorphs influencing many natural and anthropogenic processes, such as carbon sequestration. Characterizing the dissolution rate of ACC in presence of additives of contaminants in favor of crystalline phases is challenging as such reactions occur readily in bulk solution. Droplet microfluidics offers a solution by confining ACC within a droplet, enabling a quantification of the transformation rate of ACC into crystalline phases. However, accurate quantification of this transformation requires analyzing more than thousands of droplets identifying the different polymorphs of calcium carbonate during an experiment, which is labor-intensive. Here we develop a visual-based machine learning method, combining cascading U-Net and K-Means clustering, to allow efficient analysis of droplet microfluidics experiment results. Using our method, we accurately inspect 11,288 droplets over 6 hours of experimental time to identify the polymorphs, using a CPU core in a laptop for only 42 minutes. This is achieved with manual labeling of 11 experimental microscopy images before augmentations. From our analyses the transformation rate of ACC into its crystalline phases can be inferred. The transformation rate indicates an increasing stability of the ACC phase in confinement. Our method is generalizable and can be applied to different setups of droplet microfluidics experiments, facilitating efficient experimentation and analysis of complex crystallization processes.