The sample preparation has been described by . Specimens of the thickness of around 1 cm are cut with a band saw (see cutting plan in Fig. d). Both surfaces are polished with a microtome knife and exposed to air for a few hours. Sublimation smoothens the surface and creates grooves at sites of high energy, where grain and subgrain boundaries intersect the surface. In this way 2-D maps of grain boundary networks and subgrain structures can be created . When focusing into the ice volume and choosing transmission light mode, surface features such as etching grooves fade out but other objects inside the sample come into focus. Typical features are gas inclusions (air bubbles and clathrate hydrates), relaxation features (secondary bubbles, plate-like inclusions) and µ inclusions (see examples in Figs. b, c and c, d). µ inclusions appear as dark spots near the resolution limit of the microscope.

We use an optical microscope (Leica DMLM) with a CCD camera (Hamamatsu C5405), frame grabber and a software-controlled x–y stage. Ice samples of up to 10 cm side length can be scanned at the resolution of 3 µmpix-1. The final microstructure map is stitched from up to 1500 individually captured photomicrographs. We combine two of these scans for each sample – a microstructure map focused onto the surface to reveal grain boundaries and an impurity map focused ca. 300 µm into the sample volume to visualize µ inclusions (see Fig. a). Such a stack of maps allows us to study inclusions in direct relation to the grain boundary network. Since µ inclusions are mapped within the sample volume, contamination is not an issue. Due to the obliquity and 3-D shape of grain boundaries, their positions inside the sample are not exactly the same as the etching grooves on the surface. To keep this uncertainty low the second scan must be focused close below the surface. Black dots in impurity maps were detected manually. The low contrast and size of black dots in the images did not support the application of automatic detection filters. Visual detection and manual counting of µ inclusions is a time-consuming process which requires a certain amount of patience and discipline. Since it is based on observer's subjective judgment, the results may contain an observer-dependent variance. The impurity maps presented in this study were generated by three observers independently, but using the same criteria. Partial comparison of the results at overlapping regions showed a good consistency within the data and thus confirmed the observer-dependent factor being minimal.

A confocal Raman microscope was used to analyze the composition and mineralogy of selected µ inclusions. The AWI cryo-Raman system consists of a WITec alpha 300 M+ with an UHTS 300 spectrometer and a Nd:YAG laser (532 nm) set up in the cryolab at -15 ∘C . Within the scope of this paper, the measured Raman spectra shall serve as a proof that the mapped black dots are in fact chemical impurities in the form of µ inclusions. A detailed analysis of the detected minerals is being prepared for publication.

A limited number of µ inclusions (20–40 per sample) were selected for the Raman analysis. Using a 50× lens and confocal mode, a great majority (around 90 %) of the µ inclusions could be found again in the Raman microscope. Additionally, their distance from the sample surface (z coordinate) could be measured due to the confocal setting. The z coordinates of µ inclusions are Gaussian-distributed around the focal plane with a half maximum width of 200 µm. We accept this value as the depth of field of the mapping microscope and use it for the calculation of the volume fraction of the impurity maps.

Around 70 % of analyzed µ inclusions showed a Raman spectrum of sufficiently high intensity to separate it from the overall present ice spectrum. The quality of obtained Raman signal depends on several factors. The most limiting ones are the size of the µ inclusion, the path length of the laser beam through the crystal, the quality of polished surface and the acquisition time. Due to time constraint given the large amount of measured points, integration time was in the range 5–10 s with 10 repetitions. Most of the measured inclusions in EDML-2371.4 and EDML-2371.9 – around 40 spectra – could be identified as sulfate salts (see example spectrum in Fig. e). The NEEM samples showed a higher content of water-insoluble inclusions, such as quartz and black carbon. A detailed description of the composition statistics is still in progress and will be shown elsewhere. However, the Raman measurements show that the counted black dots are chemical impurities and thus verify microstructure mapping as a valid method to map the visible in situ impurity content.

The number of counted µ inclusions per sample are shown in Table . Knowing the depth of field being 200 µm we calculated the volume fraction of the mapped area. In this way, the average concentration of µ inclusions per ml of water equivalent could be estimated. Comparison with the CFA dust concentration shows a clear difference between the deep EDML ice and the shallow NEEM ice. In EDML, the number of visible µ inclusions in situ is 2–3 times higher than the content of insoluble particles (dust) in the meltwater. In contrast the amount of µ inclusion in the NEEM samples is comparable or even less than the CFA dust concentration.

We used a microscopic method to map visible impurities within the ice sample volume. Since µ inclusions are mapped in situ, surface contamination of the sample does not affect the results. The method is limited mainly by the optics of our system, in particular contrast and resolution. Objects smaller than a certain size limit would be virtually not resolved and thus a fraction of small-sized µ inclusions would be excluded from the analysis. analyzed size distributions of dust particles in the EDML ice core using a laser particle detector. The average particle diameter varied between 2 and 3 µm and only a small fraction of dust particles were smaller than 1 µm in diameter. Thus, the majority of microparticles were indeed within the resolution range of an optical microscope and our results should be comparable to the CFA dust data. Individual µ inclusions were selected for Raman measurements. The obtained spectra confirmed that the optically detected “black dots” are chemical impurities, mainly salt and dust particles. This supports previous studies by and who used Raman microscopy to analyze µ particles in ice from Dome Fuji. Detailed quantitative studies on composition and size of µ inclusions are ongoing to estimate the actual proportion of substances present as inclusions or dissolved within the ice lattice respectively.

No signal other than ice spectrum could be detected when focusing into grain interiors, grain boundaries or triple junctions. In all four samples presented in our study, impurity spectra could be detected only when focusing onto visible µ inclusions. Thus we cannot confirm observations by and , who measured acidic environments in triple junctions and grain boundaries via Raman spectroscopy, nor EDX analyses by e.g., and , who also found trace elements in grain boundaries. The AWI cryo-Raman should be considered one of the most powerful Raman systems applied to ice. Still, its spatial resolution and sensitivity are limited by the applied optics and by the physics of Raman scattering. Thus, we cannot rule out segregation of trace elements to grain boundaries if their concentrations were very low.

High-resolution CFA dust concentrations measured by were taken as reference for comparison with our concentrations of µ inclusions (Table ). The ratios dust vs. µ inclusions differ significantly between the two ice cores. The deep EDML ice in solid state contains 2–3 times more µ inclusions per volume unit than dust in the meltwater CFA;. In contrast, the NEEM samples contain comparable amounts of µ inclusions and dust. However, the NEEM CFA data still need to be flux-calibrated (Wegner, personal communication). Additionally, the presence of air bubbles in the shallower NEEM samples may cause an underestimation of µ-inclusion concentration. As air inclusions appear dark in the impurity maps (Fig. ), they may cover a substantial part of µ inclusions in the image. Another possible explanation is to assume that the NEEM samples contain predominantly insoluble dust particles which are detected in the CFA, whereas the EDML µ inclusions consist mainly of water-soluble substances. The preliminary analysis of the collected Raman spectra points in this direction. While the two Antarctic samples feature primarily sulfate salts in agreement with, we mainly found terrestrial minerals and black carbon in the Greenland samples.

µ inclusions are distributed in horizontal bands or layers of higher concentration. The annual layer thickness at 740 m in the NEEM ice core was estimated as 12–16 cm . Thus at most one annual layer would fit in one section and the layering of µ inclusions in NEEM-1346.2 and NEEM-1346.5 is probably attributed to seasonal variability. The annual layer thickness of the EDML samples (2371 m) is only 5–10 mm so that more than 10 annual layers should be present in each sample. However, the stratigraphy at this depth of the EDML ice core is strongly disrupted, strata are tilted by up to 30∘ and centimeter-sized z folds are present . The visual stratigraphy, CFA profile and c-axis orientations along the whole EDML bag 2371 are plotted in Fig. . Impurity layers are visible in the line scanner image and correlate with peaks in the dust and Ca2+ record. A sharp impurity peak is visible in the EDML-2371.9 section, which corresponds to the double horizon of µ inclusions shown in Figs. b and . The DEP and CFA conductivities show significant differences, as DEP is recorded on the ice core in solid state while CFA measures the electrolytic conductivity of the meltwater. The DEP variability is independent on dust and Ca2+ concentration, while the CFA conductivity shows a sharp peak in EDML-2371.9 correlated to the dust and Ca2+ horizon. This supports our conclusion from the last paragraph stating that a large portion of the µ inclusions in the EDML samples are water-soluble salts, which will increase the CFA conductivity. Another dust peak arises in sample EDML-2371.4, but no correlated signal is found in Ca2+ nor CFA conductivity. Furthermore, it is unclear whether the NH4+ peak in EDML-2371.4 is correlated to the mentioned dust peak and the shift is due to inaccuracy in depth assignment or the signals are independent. Further evaluation of the Raman spectra, which will be presented elsewhere, should help to better understand the origin of the signals.

The attractive force between a grain boundary and µ inclusions results from the reduction of grain boundary energy. Assuming a random distribution of spherical inclusions of radius r , the maximal pinning pressure on a grain boundary can be derived: PZ=2πr2γNV, where γ is the grain boundary energy and NV number of µ inclusions per volume unit. If the driving force for grain boundary migration PGBM<PZ, then the grain boundary stays in contact with the pinning particles and its migration rate adapts to the mobility of the particles (slow-mode pinning). In contrast, if PGBM>PZ, the interaction time is short and grain boundary proceeds with its motion, leaving the particles behind (fast-mode pinning).

One of the objectives of our study was to catch µ inclusions “in flagrante” – i.e., in the very act of pinning a grain boundary. However, this aim resulted to be cumbersome since grain boundaries are usually invisible inside the sample volume. In general, we can only estimate positions of grain boundaries from the surface images (microstructure maps). Sometimes, if the curvature and convexity are favorable with respect to the image plane, 3-D shapes of grain boundaries are indeed visible within the sample volume. In such cases, we could observe clathrate hydrates sticking to grain boundary interfaces, deforming their shapes due to the pinning force. However, no µ inclusions were observed to produce such kind of effects. We studied the distribution of µ inclusions in a plane of focus over the whole sample area. The density of µ inclusions is inhomogeneous but exhibits no correlation with grain boundaries. The typically clustered distribution could in principle be interpreted as caused by sudden release of µ inclusions from accelerating grain boundary. However, the clusters which are within themselves unordered also are of no significant shape, e.g., rows, planes or ellipsoids. We would expect some kind of alignment in rows or planes or at least some sort of graduation (size, type) of the clusters if they would result from release by a moving grain boundary.

The observations listed above lead us to the conclusion that only fast-mode pinning can take place in all four analyzed samples: PGBM>PZ. This is in agreement with , and others, who suggest that particle concentrations in ice are in general too low to induce slow-mode pinning. However, we cannot confirm the assumption by that µ inclusions would accumulate in high concentrations at grain boundaries. Our results also confirm general observations by and , who concluded that impurity layering and thus climatic record would stay preserved even in the deep parts of the EDML ice core and the NGRIP ice core respectively. identified annual layering of impurities down to the Eemian part of the NGRIP ice core. observed no redistribution of µ inclusions due to pinning or dragging down to 2300 m of the EDML ice core. However, below this depth, and particularly in the deepest 200 m, the authors found accumulations of “black dots” along grain boundaries and concluded that pinning and dragging is relevant only in the deepest part of the EDML ice core. This finding mismatched our observations in EDML-2371.4 and EDML-2371.9, where we only found secondary bubbles accumulated along grain boundaries. Since microstructure maps described by were made only few days or hours after drilling the ice core, we decided to record new microstructure maps of the same spots now after ca. 10 years of storage. The comparison indicates that significant changes occur during the relaxation of the material (see example in Fig. ). The image of the freshly drilled ice shows a group of “black dots” accumulated at a grain boundary. The image of a re-measurement of the same sample shows that with time these “black dots” have grown and filled with gas and now they are forming typical relaxation air bubbles. There are two possible explanations. (1) Grain boundaries in deep EDML do collect µ inclusions by dragging but after retrieving the core they serve as seeds for the growing relaxation bubbles and get surrounded with gas. It remains unclear why µ inclusions in the interior of grains do not evolve into bubbles. (2) Another option is that the “black dots” at grain boundaries observed by in fact are small micro-bubbles forming due the abrupt drop of pressure after logging the ice core. analyzed compositions of secondary microbubbles using Raman spectroscopy. The authors also remark that secondary bubbles tend to form at locations of “black dots”. However, to date there are no data available concerning the composition of these original “black dots” as Raman spectroscopy is currently not available on-site, viz. right after drilling.

observed impurity concentration peaks at grain boundaries in the clean parts of the Greenland Stadial 22, NGRIP ice core, using LA-ICPMS. However, fundamental differences between the experimental techniques as well as the ice samples impede a direct comparison to our study. Dust concentrations in GS-22 determined by vary between 2×105 mL-1 in cloudy bands and 2×104 mL-1 in the clean bands, which is 1 order of magnitude more than the average concentrations in our samples. Furthermore, it remains unclear what portion of the LA-ICPMS signal is due to visible µ inclusions. A comparative study applying both techniques to the same specimen could help resolve the contradiction.

The highest local µ-inclusion density in our sample material was found in the double horizon of sample EDML-2371.9 (Figs. , ) and could be estimated as NV=37206 cm-3. When inserting this value in Eq. (1), assuming the mean particle radius r=1.5 µm in agreement with and using γ=65 mJm-2 for high angle boundaries , we obtain the maximal pinning pressure PZ=0.034 Nm-2. The driving force for grain boundary migration can be written as PGBM=ΔH-2γR, where ΔH is the gradient in stored strain energy density across the grain boundary (i.e., strain-induced driving pressure) and the second term represents the curvature-driven pressure with the radius R of grain boundary local curvature. Assuming no difference in stored strain energy (ΔH=0), the boundary migration will be driven only by its curvature. Inserting PZ and Eq. (3) into Eq. (2) we obtain R<3.8 m the radius of local curvature necessary to unpin the grain boundary from its surrounding µ inclusions. Most grain boundaries along the NEEM and EDML ice cores indeed fulfill this condition e.g.,.

In a second marginal case, let us consider a planar grain boundary (R→∞) whose migration is only driven by the difference in stored strain energy as function of the local dislocation density Δρdis. Following and , ΔH can be approximated as ΔH=Δρdis12Gb2, where Gb2/2 is the energy per unit length of a single dislocation line consisting of the shear modulus G and the magnitude of the Burgers vector b. Considering only the basal slip system the mean dislocation energy can be estimated as Gb2/2=3.6×10-10 Jm-1. Inserting this in Eq. (4) and combining Eqs. (4), (3) and (2) we obtain an estimation for the minimal difference in dislocation density required for unpinning: Δρdis>108 m-2. This value is 2–3 orders of magnitude smaller than absolute dislocation densities modeled by and and also smaller than a minimum dislocation density access calculated from grain boundary curvatures by Fig. 11b.

In the above considerations we made a variety of assumptions and the derived values are only for approximation. Furthermore, in the case of a real grain boundary ΔH acts against the curvature-driven force as described in Eq. (3) and thus the final motion and shape are determined by the ratio of these forces. However, the exercise demonstrated that applying Zener's theory to our particular µ-inclusion concentration the pinning effect is comparatively small and will hardly affect grain boundary migration. This is indeed in agreement with what we observe.

It is difficult to make general conclusions based on the analysis of four discrete samples. Our study indicates that pinning on µ inclusion occurs in fast mode in a large part of the ice sheets, as already stated by , and will not significantly affect grain boundary migration. At the same time, with our Raman microscope we find no other form of, for example, dissolved impurities segregated to grain boundaries. However, negative correlations between average grain size and impurity content were found in virtually all ice cores e.g.,. High-impurity ice exhibits generally smaller grains than low-impurity ice at the same depth. This negative correlation can be found at all scales: in seasonal variabilities (cloudy bands, cm), during rapid climatic fluctuations such as Dansgaard–Oeschger events (tens of meters) or comparing glacial and interglacial periods (hundreds of meters).

Grain size has an impact on a variety of ice physical properties ; vice versa it is controlled by thermodynamic conditions and processes within the ice sheet. Without deformation and under purely static conditions, mean grain area increases linearly with time. This NGG is driven by the reduction of grain boundary surface energy due to the optimization of volume versus interfaces. This model is especially relevant for very small grain sizes well below the equilibrium or steady-state grain size see, e.g.,, such as smallest grain sizes in the uppermost part of the ice sheet , but may also apply to some areas where topographic depressions in the bedrock below ice sheets inhibit deformation and NGG can produce extraordinary large grain sizes stagnant ice;. If deformation introduces additional energy into the system, dynamic recrystallization processes are activated, driven by the energy reduction. Rotation recrystallization (RRX) splits grains into subgrains and thus has a grain-size-reducing effect . During RRX recovery orders dislocations of a deformed grain into subgrain boundaries, which are lower energy states of dislocation assemblages . Subgrain boundaries can develop into grain boundaries by further rotation. In contrast, strain-induced boundary migration (SIBM) occurs when grain boundaries propagate into regions of high dislocation densities. The effectiveness of SIBM increases with the heterogenous distribution of dislocations in polycrystalline ice, which is a direct consequence of the high mechanical anisotropy of the ice crystal. SIBM leads to huge grain sizes in the deep ice but can also lead to grain size reduction, e.g., by nucleation of new grains SIBM-N; or by dissection of highly irregular grains. Recent microstructural studies e.g., show that all recrystallization mechanisms (NGG, RRX and SIBM) concur all over the depth range of an ice sheet rather than being dominantly active in separate depth zones. The grain size is then a product of the interplay between these processes dynamic grain growth;.

It is widely accepted that the grain size is modulated by some impurity effect. The two most manifest candidates for such an effect are (1) reduction of grain boundary mobility via dissolved impurities being dragged along by migrating grain boundaries ; (2) Zener pinning , that is, interaction between µ inclusions and grain boundaries as discussed in Sect. . However, experimental evidences for both models are controversial in ice, as demonstrated by our study. Furthermore, both interpretations are based on the assumption that it is solely NGG, which suffers under the effect of impurities leading to smaller grain sizes. However, grain size is a product of all recrystallization processes together, as discussed previously. Therefore we hypothesize that an indirect impurity effect, for instance enhanced deformation and/or effect on dynamic recrystallization triggered by high impurity content, could be an alternative candidate responsible for the changes in grain size. Impurities could have a significant influence on strain distribution within grains as well as dislocation mobilities and densities, e.g., via dislocation multiplication. The microstructure effects of increasing dynamic recrystallization versus viscoplastic deformation have recently been tested by a microstructural evolution model (Llorens, 2016a, b). In these model runs the impact is significant on, e.g., the grain shape evolution because deformation tends to flatten grains while recrystallization tends to make them equidimensional. First evidences of changes in the deformation–recrystallization interplay have been observed by means of grain shape analyses . However, to test these hypotheses is far beyond the scope of this study.