The freezing of supercooled water to ice and the materials which catalyse this process are of fundamental interest to a wide range of fields. At present, our ability to control, predict or monitor ice formation processes is poor. The isolation and characterisation of frozen droplets from supercooled liquid droplets would provide a means of improving our understanding and control of these processes. Here, we have developed a microfluidic platform for the continuous flow separation of frozen from unfrozen picolitre droplets based on differences in their density, thus allowing the sorting of ice crystals and supercooled water droplets into different outlet channels with 94 ± 2% efficiency. This will, in future, facilitate downstream or off-chip processing of the frozen and unfrozen populations, which could include the analysis and characterisation of ice-active materials or the selection of droplets with a particular ice-nucleating activity.