Direct visualization of solute locations in laboratory ice samples

Many important chemical reactions occur in polar snow, where solutes may be present in several reservoirs, including at the air–ice interface and in liquid-like regions within the ice matrix. Some recent laboratory studies suggest chemical reaction rates may differ in these two reservoirs. While investigations have examined where solutes are found in natural snow and ice, few studies have examined either solute locations in laboratory samples or the possible factors controlling solute segregation. To address this, we used micro-computed tomography (microCT) to examine solute locations in ice samples prepared from either aqueous cesium chloride (CsCl) or rose bengal solutions that were frozen using several different methods. Samples frozen in a laboratory freezer had the largest liquid-like inclusions and air bubbles, while samples frozen in a custom freeze chamber had somewhat smaller air bubbles and inclusions; in contrast, samples frozen in liquid nitrogen showed much smaller concentrated inclusions and air bubbles, only slightly larger than the resolution limit of our images (∼ 2 μm). Freezing solutions in plastic vs. glass vials had significant impacts on the sample structure, perhaps because the poor heat conductivity of plastic vials changes how heat is removed from the sample as it cools. Similarly, the choice of solute had a significant impact on sample structure, with rose bengal solutions yielding smaller inclusions and air bubbles compared to CsCl solutions frozen using the same method. Additional experiments using higher-resolution imaging of an ice sample show that CsCl moves in a thermal gradient, supporting the idea that the solutes in ice are present in mobile liquid-like regions. Our work shows that the structure of laboratory ice samples, including the location of solutes, is sensitive to the freezing method, sample container, and solute characteristics, requiring careful experimental design and interpretation of results.


37
A variety of potential chemical reactants have been identified in snowpacks; concentrations can 38 vary considerably, with typical concentrations on the order of 10 µM in clean Arctic snows (Yang et al., 39 1996). Impurities can integrate into snow crystals during formation, or be deposited onto the surface of 40 formed crystals. Reactants and products also partition between the snow crystals and the overlying air; 41 the large surface area of the snow crystals provides an extensive environment for reactions to occur. As 42 the snowpack consolidates and snow grains metamorphose, chemical compounds can remain at the 43 surface of the crystals, or become trapped internally at grain boundaries or triple junctions (  Rausch et al., 2014; Domine et al., 2008;Grannas et al., 2007). 45 There appear to be three reservoirs for impurities in snow: a quasi-liquid layer (QLL) at the ice-46 air interface; liquid-like regions (LLRs) within the ice (e.g., at grain boundaries); and in the bulk ice the location is important for several reasons. First, chemicals in a surface QLL can be more readily 50 released to the atmosphere compared to impurities segregated into an internal LLR; furthermore, gas-51 phase oxidants and other species can readily partition from the air onto solutes at the air-ice interface. 52 Second, photon fluxes can vary considerably in various locations within the snowpack (Phillips and 53 Simpson, 2005), although there appear to be only small differences within crystals themselves (McFall 54 and Anastasio, 2016). Third, the rates of reactions of impurities appear to vary with location. For 55 example, photolysis rates of PAHs (polycyclic aromatic hydrocarbons) have been reported to be up to five 56 times faster in surface QLLs compared to in whole ice samples (where PAHs are likely in LLRs) or in 57 aqueous solution Donaldson, 2007, 2010; Ram and Anastasio, 2009). An investigation of 58 reactions in frozen solutions (Kurkova et al., 2011) suggested the QLL and LLR physical reaction 59 environments are substantially different, with QLLs best represented by a 2D cage and LLRs as a 3D 60 cage. This work also found that the cage effect (i.e., the tendency for a compound to be surrounded by 61 solvent molecules, which can impede the ability of a compound to react) at a given temperature was much 62 more pronounced for reactions occurring in QLLs than LLRs, with solutes in QLLs having less mobility 63 compared to solutes in LLRs.

64
Because of the potential reactivity differences between the reservoirs, understanding reaction 65 rates in different reservoirs requires knowing where solutes are located.  Yang et al., 1996). The chosen concentration allows easy visualization in our system and provides 106 enough material to evaluate spatial patterns in the sample.

107
We froze most samples as a 500 µl aliquot in a capped glass vial (approximately 3 cm high and 1 108 cm in diameter, 0.8 mm wall thickness, with a total vial volume of ~2 ml) using one of three methods. 109 These methods were chosen because they had been used in our laboratory, as well as others, and also due 110 to differences in the speed of heat removal from the samples; we discuss later the expected morphologies 111 for the various sample types. In the first technique ("Freezer"), we placed samples upright on a plastic 112 plate in a laboratory freezer at approximately -20° C; freezing took approximately 1 hour. In the second 113 technique ("Freeze Chamber"), we froze samples upright in a custom-built freeze chamber (Hullar and 114 Anastasio, 2011) whose base was cooled to either -10 or -20° C. Typically, the sample sat directly on the 115 base of the freeze chamber surrounded by air. However, we also froze some samples surrounded by 116 drilled metal plates, effectively placing the sample in a metal "well"; the distance between the sample and 117 the surrounding plates was around 1 mm. In the third technique ("Liquid Nitrogen" or "LN2") we froze 118 samples by putting the aqueous sample in a vial, capping it, then immersing it in a bath of liquid nitrogen 119 deeper than the height of the liquid in the vial; freezing time was ~30 seconds. We allowed all samples to 120 anneal at -10° C for at least 1 hour before imaging. We froze a small number of samples in either 121 polypropylene vials (wall thickness ~1 mm) or with a larger sample volume (750 µl).

122
We imaged samples using a MicroXCT-200 (Zeiss Instruments) micro-computed tomography 123 (microCT) scanner. To maintain our samples at -10° C , samples were held in a custom cold stage for the 124 MicroXCT-200 (Hullar et al., 2014). The custom cold stage was placed on the scanner's sample stage, 125 whose position is controlled by the scanner software to submicron precision. Scanning parameters were 126 set based on the manufacturer's guidelines. For most imaging, we set source and detector distances to 40 127 and 130 mm respectively; voltage and power were set at 70 keV and 7.9 W, and the manufacturer's LE3 128 custom filter was used for beam filtration. The microCT acquired 1600 projections over 360 degrees of 129 rotation, with an exposure time of 2 s. Images were reconstructed using the manufacturer's software on 130 an isotropic voxel grid with 15.9358 µm edge lengths. Some samples were analyzed at higher resolution, 131 with a voxel edge length of 2.1146 µm. For these samples, we set source and detector distances to 60 and 132 18 mm, used the LE5 beam filter, collected 2400 projections spanning 360 degrees, and set beam voltage, 133 power, and exposure time to 60 keV, 6 W, and 30 s respectively. The microCT scanner software outputs 134 slicewise TIFF images of the X-Y plane of the sample, with greyscale values corresponding to the 135 radiodensity of each voxel at that Z plane.

136
We imported digital TIFF images into the Amira software package (Visualization Sciences 137 Group, FEI) for reconstruction and segmentation. Our segmentation procedure used the Amira 138 segmentation tools to isolate the sample from surrounding materials; generally, our procedure should 139 include very little sample container at the expense of excluding some small amounts of sample in contact 140 with the vial wall. Similarly, the segmentation procedure excludes very little sample in contact with air 141 above the sample, while including small amounts of top air as sample. Some images presented here were 142 mathematically smoothed by the software, which sometimes resulted in small features (< 80 µm in 143 diameter) being eliminated from movies and still images; however, smoothing did not substantially 144 change the interpretation of our results. In some cases we prepared histograms of the data, which were not 145 smoothed and include all sample data. 146 To quantitate CsCl concentration in each voxel, we first imaged samples of Milli-Q water, as both 147 liquid and ice, and measured the average radiodensity (image greyscale value) of a subvolume within 148 each sample. As expected, the average radiodensity of ice (4948 ± 160 (1σ)) was less than that of liquid 149 water (5372 ± 194 (1σ)) due to the lower density of ice. Our measured radiodensity ratio between ice (at -150 10° C) and water (at 20° C) was 0.921, matching a calculated density ratio from literature values (Haynes,151 2014) of 0.921. Next, we imaged 8 aqueous solutions of CsCl at varying concentrations (1.0 mM to 5.0 152 M) to construct a calibration curve. Plotting these points (Fig. 1) shows a linear relationship between 153 CsCl concentration and measured radiodensity, with a y-intercept value within the range of our measured 154 radiodensities for pure liquid water. Therefore, the measured radiodensity of a voxel within a sample 155 containing CsCl in solution (or ice) is linearly related to the amount of CsCl present in the voxel. We 156 assume the relationship between CsCl concentration and radiodensity is the same for ice and water. This 157 allows us to determine the amount of CsCl present in a sample voxel by subtracting the average greyscale 158 value of pure water (or ice) and then using the standard curve to calculate the CsCl mass. 159 When aqueous solutions are frozen, solutes are generally excluded from the forming ice matrix, 160 resulting in a two distinct components: pure (or nearly pure) water ice, and a concentrated solution of LLR concentration is considerably lower than the solubility limit of CsCl (11.1 M at 20° C, 9.6 M at 0° C 167 (NIH, 2015)), but higher than the solubility limit of Rose Bengal (1 mM, temperature not given, (Neckers,168 1989)). Therefore, we do not expect CsCl to precipitate, although Rose Bengal might. 169 As described earlier, we use the LLRs would be somewhere between 3 and 3.2 M, i.e., 10 -20% higher than our ideal case concentration, 189 but neither source presents freezing point depression data measured at such a high concentration. and voxels containing a substantial amount of solute (V LLR /V VOXEL > 10%). We define an "air" voxel as 198 having a radiodensity less than or equal to the average radiodensity of an imaged air sample, i.e., 3996. 199 As noted above, greyscale values from images of pure materials vary somewhat, meaning a clear 200 distinction between two materials with similar average greyscale values is not possible. We chose to set 201 the cutoff for segmenting LLRs at a greyscale value of 5507, a threshold three standard deviations greater 202 than the average greyscale value for pure ice, which will essentially eliminate the problem of identifying 203 water ice as solute. However, because of this high threshold it is quite likely that solute is present in some 204 voxels characterized as "ice". On the other hand, voxels defined as having an LLR percentage of 2% or 205 greater almost certainly contain solute. For CsCl-containing samples, we calculated the mass of CsCl 206 present in each domain. Because the statistical distributions of voxels containing only pure water ice and 207 those containing <2% LLR as well as pure water ice overlapped , we could not determine the mass of 208 CsCl present in the "ice" domain directly. Therefore, we assumed any mass not present in either the LLR 209 2-10% or LLR >10% domains is present in the "ice" domain. 210 211

Results and Discussion 212
We first present imaging results for samples prepared without added solute (frozen Milli-Q 213 water). Figure 2a shows a reconstructed image of a "pure" ice sample prepared by freezing air-saturated 214 Milli-Q in a glass vial in a laboratory freezer; the full movie, which shows the sample rotating, is in 215 Supplemental Fig. S1. Air bubbles are visible as light grey spheroids, and are generally located towards 216 the center of the sample, away from the vial walls and base. This is likely because the entire outer surface 217 of the vial was cooled and the water apparently froze from the outside inward. Supporting this idea, some 218 of the bubbles appear to elongate along the radial axes of the sample, similar to the bubble elongation 219 seen by Carte (Carte, 1961) in a temperature gradient. The isolation of bubbles within the middle of the 220 sample seems to follow Shumskii's (Shumskii, 1964) model of the formation of the "central nucleus", 221 with impurities (in this case, air bubbles) forced to the center of a freezing water mass. 222 Figure 2b shows a reconstruction of a similar Milli-Q sample, but now where the solution was 223 degassed with helium for 30 minutes before freezing; the full movie is in Supplemental Fig. S2. Because 224 He degassing replaces the more soluble nitrogen and oxygen in the air-saturated solution with less soluble 225 helium, fewer bubbles are present in Fig. 2b. The size of the bubbles, however, is roughly similar in the 226 two figures (approximately 150-300 µm), suggesting bubble size is a function of the freezing method, not 227 of the gas itself. 228 Figure 2c shows a histogram of the number of voxels containing various radiodensities, 229 represented here as the ratio V LLR /V VOXEL , in the two water ice samples. A ratio of zero represents the 230 average radiometric density for pure water ice, with values slightly greater or less than zero indicating 231 noise in the sample images and reconstruction. Voxels containing only air comprise the smaller, second 232 peak centered at approximately V LLR /V VOXEL = -0.05, which overlaps with the primary (pure ice) peak.

233
Taking into account that the Y axis (voxel count) is a log scale, the two curves show the volume of gas 234 bubbles is clearly less for the helium degassed treatment. Table 1 shows the estimated volumes of water 235 ice and gas bubbles in the two samples, as determined by our segmentation process (see the Methods 236 section). The gas volume in ice made from air-saturated water is approximately 1.4 %, while the ice 237 made from helium-saturated Milli-Q has approximately half the gas volume. Figures 2a and 2b appear to  238 show a larger difference in gas volume between the two samples, suggesting that many of the small 239 bubbles in the sample imaged in Fig. 2b  Next, we examined the effect of freezing method on both freezing morphology and solute 245 location. The Freezer, Freeze Chamber, and LN2 sample preparation methods are described in the 246 Methods section. Figure 3 shows the results of imaging several combinations of freezing method and 247 solute. We start with an image of the ice made by freezing 1.0 mM CsCl in a laboratory freezer. As 248 shown in Fig. 3a (and the Supplemental Fig. S3 movie), both air bubbles and concentrated CsCl LLRs are 249 relatively large, with the LLRs tending to wrap around the air bubbles. Figure 3b is a magnification of 250 the red-bordered area in Fig. 3a, showing examples of large solute inclusions wrapped around air bubbles 251 (lighter gray spheroids). 252 Figure 3c (movie: Supplemental Fig. S4) shows a similarly prepared sample as the Freezer 253 sample in Fig. 3a, but frozen in our Freeze Chamber. Compared to the Freezer sample, the Freeze 254 Chamber sample has smaller air bubbles and inclusions, more solute present near the top of the sample, 255 and the areas of concentrated solutes (LLRs) are less likely to be associated with the air bubbles. These 256 points are clearly shown in Fig. 3d, which is a magnification of the red bordered area of Fig. 3c. 257 Considering that these two samples were frozen at similar temperatures, the morphologies are 258 substantially different. As seen in Table 1, the fraction of voxels containing a LLR fraction >10% is 259 about five-fold less in the Freeze Chamber sample than the Freezer sample, while the fraction of voxels 260 with an LLR concentration between 2 and 10% doubles. This finding indicates the freezing process in the 261 freeze chamber creates smaller LLR inclusions than does the freezer, with LLRs distributed more widely 262 throughout the sample. Additionally, substantial amounts of solute were segregated towards the surface 263 of the Freeze Chamber sample; presumably, the sample froze from the bottom and solutes were 264 preferentially excluded from the advancing freezing front. However, the same process did not affect the 265 air bubbles, which are well distributed throughout the sample. We believe these structural differences may 266 be due to faster freezing in the Freeze Chamber sample, as the freeze chamber removes heat more quickly 267 than the freezer because of direct contact between the bottom of the vial and the chilled base plate in the 268 chamber. Previous work (Hallett, 1964;Rohatgi and Adams, 1967) has shown faster freezing gives closer 269 spacing of ice dendrites or plates in the sample as it freezes, which then leads to smaller solute inclusions 270 or bubbles, similar to our finding here. Supplemental Fig. S5 shows a sample prepared in the same way 271 as in Fig. 3c, but with the metal plates in place in the freeze chamber, which surrounds the vial with metal (concentrated solute) and dark (air bubble) areas, suggesting some segregation of CsCl and air occurs 288 even with rapid freezing (~30 seconds). However, this effect is less noticeable in the quickly frozen 289 liquid nitrogen sample (Supplemental Fig. S7), and much more pronounced in the other two freezing 290 methods (Figs. 3a and 3c). Analogous findings, although using a very different experimental system, 291 were reported by Heger et al. (Heger et al., 2005), who found solutes were concentrated by as many as six 292 orders of magnitude with slow (several minutes) freezing, but only three orders of magnitude when frozen 293 in liquid nitrogen. 294 Figure 3g shows the histogram for the 1.0 mM CsCl solutions frozen using each of the three 295 freezing methods, as well as for Milli-Q water ice frozen in a laboratory freezer. Unlike the images seen 296 in Figs. 3a through 3f, where mathematical smoothing can eliminate small structures, the histograms 297 include all the voxels in the sample. As discussed in Fig 2c, water ice has two overlapping peaks, 298 corresponding to air bubbles (left peak) and ice (right peak). Some voxels, shown in the "saddle" 299 between the two peaks, contain both air bubbles and pure water ice, and will therefore have a greyscale 300 value between air and ice. The Fig. 3g  inclusions in the bulk ice, the calculated ratio in a voxel could be higher than 1. The fact that the ratio 305 gets close to, but never exceeds, 1 is consistent with our tricomponent model of air, relatively pure ice, 306 and concentrated LLRs with a maximum concentration of 5.4 M total solute. 307 The increased number of air voxels on the left end of the curve for the 1.0 mM CsCl freezer 308 sample represents voxels composed entirely of air. This number is larger than in the water sample, 309 supporting the imaging findings that the presence of solute actually increases the size of air bubbles. For 310 the Freeze Chamber and LN2 samples, the number of voxels containing only air is smaller, and voxels 311 containing air are more likely to contain at least some fraction of ice or solute. For the Freeze Chamber 312 sample, the histogram correlates with the images (2c and 2d), with fewer voxels containing a large 313 volume fraction of highly concentrated regions than in the Freezer sample. Finally, the liquid nitrogen 314 histogram is nearly identical to water ice, although a few voxels with concentrated solute are present (also 315 seen in Supplemental Fig. S7). Next, we examined the impact of solute on freezing morphology and 316 solute location, by replacing CsCl with Rose Bengal, a large, organic molecule (see structure in 317 Supplemental Fig. S8). Figure 3f (movie: Supplemental Fig. S9) shows a sample containing 1.0 mM 318 Rose Bengal frozen in our freeze chamber. Using 1.0 mM Rose Bengal instead of 1.0 mM CsCl (Fig. 3c)  319 gives a very different freezing pattern, with only a few small bubbles and no visible areas of concentrated 320 solute. While mathematical smoothing has likely eliminated some of the smaller structures, the overall 321 sample morphology is quite different than that produced by the same concentration of CsCl. Miedaner 322 and Miedaner and co-workers (Miedaner, 2007;Miedaner et al., 2007), using different compounds, also 323 found that sample morphology was highly sensitive to solute identity. Interestingly, changing solute in 324 our system alters not only the structure of solute inclusions, but also the size of the air bubbles. The exact 325 reason for the change in morphology is unclear. CsCl is more polar than Rose Bengal, and could 326 influence the movement of the polar water molecules into the forming ice matrix. As a relatively large 327 organic molecule, Rose Bengal might potentially modify the ice matrix due to its size. Finally, we note 328 the thermodynamically predicted final concentration of solute ions at -10° C is 5.4 M; at this 329 concentration CsCl should still be in solution, while a substantial portion of the Rose Bengal should have 330 precipitated. Whether precipitated Rose Bengal is present as solids incorporated into the ice matrix or as 331 precipitates in LLRs is not known. 332 The reproducibility of samples prepared on different days but using identical methods was quite 333 good, with similar patterns seen for each replicate (Supplemental Fig. S10). Each combination of 334 freezing method and solute gave a distinct distribution of solute and air bubbles, suggesting these two 335 variables have a significant impact on ice morphology in our experimental system. 336 Table 1 lists the calculated volume of each material domain and the total CsCl mass present, 337 including all sample voxels, based on segmentation described in the Methods section. As seen in the 338 images and histogram, the Freezer sample has the highest fraction (0.00019) of voxels containing 10% or 339 more LLR volume, approximately 5 times greater than the Freeze Chamber sample. In contrast, the 340 fraction of voxels with V LLR /V VOXEL = 2-10% in the Freezer sample (0.003) is about half that in the Freeze 341 Chamber sample, and the fraction of gas bubbles appears to be less than in the Freeze Chamber sample. 342 However, this may be a computational artifact; voxels containing LLR next to gas bubbles will have a 343 greyscale value somewhere between air and LLR, and therefore may be mistakenly counted as water ice 344 voxels. Unfortunately, determining the magnitude of this error is difficult -requiring estimating the 345 surface area of both air bubbles and any adjacent LLRs to identify suspect voxels -and is beyond the 346 scope of this study. Because LLRs in the Freezer samples are more concentrated and appear to be more 347 frequently found next to air bubbles (as seen in Fig. 3b), this effect may be more pronounced in the 348 Freezer samples than Freeze Chamber samples. However, the number of voxels mistakenly classified as 349 water (or less concentrated solute) is limited to boundaries between air and LLRs and therefore small, and 350 should not affect the overall interpretation of results. Examining the location of the CsCl mass, more than 351 10% of all CsCl present in the Freezer sample is found in voxels with LLRs >10%, while in the Freeze 352 Chamber sample only around 1% of the mass is found in these most concentrated LLRs. For both Freezer 353 and Freeze Chamber samples, about two-thirds of the CsCl mass is found in the ice compartment, 354 suggesting most solutes are present in very small LLR inclusions that are indistinguishable from water 355 ice. For the LN2 sample, only 12% of the mass is found in detectable LLRs, with the remainder 356 distributed throughout the water ice. It is also possible that the CsCl in the LN2 samples is present not as 357 liquid inclusions, but as solid solution within the water ice. However, the solubility of HCl in solid ice is 358 (1-2) x 10-4 M (Gross et al., 1975), while the CsCl solubility in solid ice would need to be 5-10 times 359 greater, assuming all the CsCl is present in solid solution. The "missing" CsCl mass here is 0.88 * 126.3 360 µg = 111.1 µg, or 0.66 µmol. Assuming this solute is entirely present as LLRs with solute concentration 361 of 2.7 M, this equates to a total LLR volume of 0.24 µL. The volume of pure ice (again from Table 1) is 362 725 µL. Therefore, assuming the remaining CsCl is distributed equally throughout the voxels labeled as 363 pure ice in Table 1, the calculated average V LLR /V VOXEL for these voxels is 0.034%, indistinguishable 364 from water ice in our system. While it is possible the CsCl is present (at least partially) as solutes in the 365 solid ice matrix, we believe it is more likely to be present primarily as small LLR inclusions. 366 Additionally, we present evidence later in this paper supporting the idea that solutes are predominantly 367 present as LLR inclusions. 368 We next examined the impact of sample container on sample morphology and solute distribution 369 by imaging samples frozen in plastic vials instead of the glass vials we used above. While many of the 370 samples discussed thus far were frozen in the laboratory freezer, most of the samples prepared in plastic 371 vials were frozen in the freeze chamber. Therefore, to allow appropriate comparisons, we first present a 372 sample of water (no solute) frozen in the freeze chamber and compare this with previous samples frozen 373 in the freezer. Milli-Q water frozen in the freeze chamber in a glass vial (Supplemental Fig. S11) gives 374 similar spatial distribution and somewhat smaller air bubbles sizes as a similar sample frozen in a 375 laboratory freezer ( Fig. 2a and Supplemental Fig. S1). However, freezing water in a plastic vial rather 376 than glass can make a significant difference in ice morphology, as shown in Supplemental Fig. S12. 377 While ice in a glass vial forms many roughly spherical bubbles, water frozen in a plastic vial using our 378 freeze chamber forms long vertical channels; such directional growth of air bubbles in a freezing liquid 379 has previously been reported (Carte, 1961). While the reason for this morphology is not entirely clear, , 380 we believe it is related to how heat is removed from the sample during freezing.. Because plastic 381 conducts heat more poorly than water, ice, or glass, the vial walls act as insulators, forcing heat to be 382 primarily removed from the bottom of the sample where the plastic vial contacts the chilled plate at the 383 base of the freeze chamber. This may promote the formation of vertical air channels as the ice freezes 384 upwards through the sample, rather than from the walls towards the interior in the glass vial sample. 385 We next examine the impact of freezing in plastic for a sample containing solutes. Supplemental 386 Fig. S13 shows a 1.0 mM CsCl solution frozen in the freezer in a plastic vial; compared to the similarly 387 treated sample frozen in a glass vial (Fig. 3a), the air bubbles and concentrated inclusions are smaller in 388 the plastic vial. Interestingly, the air bubbles in the plastic vial CsCl Freezer sample do not show any of 389 the elongation found when Milli-Q water is frozen in a plastic vial in the freeze chamber (Supplemental 390 Fig. S12), which may be related to the directional heat removal in the freeze chamber. Finally, once again 391 using the freeze chamber, Supplemental Fig. S14 shows 1.0 mM Rose Bengal frozen in plastic in the 392 freeze chamber. Here, we see substantial volumes of LLRs and more bubbles than seen in the sample 393 frozen in a glass vial, but without any elongation to bubbles or LLRs.

394
We also performed several other experiments to examine the nature of LLRs. Figure 4  CsCl is contained in liquid-like regions in our ice samples. We placed a 1.0 mM CsCl sample (glass vial; 407 Freezer) in the microCT sample holder set at -10 °C and took images of the sample (2 µm voxel 408 resolution, X-Z plane) at 0, 11, and 22 hours. The temperature gradient in the sample holder was 409 measured later by placing a thermocouple sensor between the glass vial and the holder wall at various 410 positions. The temperature difference between the bottom and middle of the holder (approximately 1.7 411 cm, extending above and below the 1 cm height of the frozen sample in the vial) was 2.2 °C, resulting in a 412 temperature gradient of 0.13 °C mm -1 . As seen in the three images, over the 22 hours of this experiment 413 the bright areas of CsCl move in the direction of the temperature gradient, towards the warmer top of the 414 vial, at a rate of approximately 10 µm h -1 . (i.e., 7.7 µm h -1 /(K -1 cm -1 )). In many cases, the solutes appear 415 to be migrating around the surfaces of air bubbles, which are visible as darker grey spheres. While the air 416 bubbles appear to remain stationary in the ice matrix, with an estimated maximum migration rate of 0.15 417 µm h -1 /(K -1 cm -1 ), the CsCl moves. Solutes are excluded from the forming ice matrix during freezing 418 (Hobbs, 1974;Petrenko and Whitworth, 1999); here, it appears the solutes are present as a concentrated 419 liquid-like solution, which can migrate either along the boundaries between air bubbles and the bulk ice, 420 or possibly by melting into the bulk ice itself (Notz and Worster, 2009 a "Gas" is defined as having a greyscale value of < 3996, "Water ice" is defined as containing < 2% liquid-like region (LLR), "LLR 2-10%" is 668 water ice containing an LLR fraction of between 2 and 10%, and "LLR > 10%" is water ice containing > 10% LLR. The original sample volume 669 (either 500 or 750 µL) is not fully captured in the volumes reported here. The segmentation process eliminates some of the lower part of the 670 sample, reducing the reported volume somewhat.