Experiments were conducted using a homebuilt cold stage based on a temperature-controlled aluminum block with a hole in it for light transmission (fig. S1). A polycrystalline diamond slide, which has good thermal conductivity, was fixed on the aluminum block over this hole using thermal paste to create a cold surface of uniform temperature. Samples were placed onto the diamond window using a thin layer of vacuum grease. Cooling of the aluminum block was achieved using a liquid nitrogen flow, and the temperature was controlled by counterheating with cartridge heaters embedded in the stage. The temperature was measured with a platinum resistance thermometer inserted into the aluminum block. The cold stage was mounted onto an Olympus BX53 light microscope, and a Phantom Miro 320 high-speed camera was used to collect freezing videos at 3000 fps. A Perspex cell with an inlet and an outlet for gas was placed around the sample. The bottom of this cell sat on the aluminum block, or on the edge of the thin section depending on the sample size, while the top contained a hole into which the microscope objective was lowered. A flow of zero-grade nitrogen was applied to the inlet to remove water from the atmosphere within the cell to avoid condensation of water onto the substrate as it was cooled.

Experiments were conducted by pipetting a 1-μl drop of Milli-Q water (18.2 megohm·cm) onto the selected substrate. The Perspex cell was then placed around the sample, the stage was raised to focus on the sample surface, and the drop was then cooled at a constant 1°C min−1 from 15°C until the droplet froze. The high-speed camera was operated using the Phantom Camera Control software. During operation, the camera would constantly record to its internal memory at the desired framerate and resolution. When ice nucleation was observed, the software was triggered. This stopped the recording, and the data that had been recorded to memory over the previous 1.5 s, containing the nucleation event, were stored. At the point of triggering, the temperature measurement displayed on the Eurotherm 2416 was recorded. Once a freezing event was recorded, the Eurotherm 2416 controller was used to raise the temperature to 5°C at a heating rate of 5°C min−1 to allow the drop to melt. The temperature was held until no ice remained in the drop, typically for 1 min, and then the cooling cycle was restarted at 1°C min−1. Each experiment comprised between 10 and 30 of these freeze-thaw cycles. For drops that froze at lower temperatures, a higher nitrogen flow rate was typically required to avoid freezing by communication via water condensed on the substrate. In total, nine separate experiments were performed using LD3 microcline, each on different regions of the feldspar surface. Five different thin sections were used in these experiments [three with (010) orientation and two with (001) orientation]. For rose quartz, two experiments in different locations on the surface of one thin section were performed.

The temperature uncertainty within an individual freeze-thaw experiment was estimated as 0.2°C, determined by measuring the melting points of water, dodecane and undecane, as used by Whale et al. (39). This uncertainty is consistent with that estimated by Atkinson et al. (11) using the same cold stage. Between experiments, that is, when a new area was investigated and freeze-thaw cycles restarted, the temperature uncertainty increased to ca. 1°C, due to factors such as the varying thermal conductivity of different thin sections. In all cases, these experiments were performed on glass slides of the same thickness as those used for preparing thin sections to minimize uncertainties from thermal gradients.