| 25 Jun 2026
Continuous Microstructural Characterization of a shallow East Antarctic Ice Core Using Machine Learning Super-Resolution and Micro-CT Imaging
Abstract. This study presents a continuous, high-resolution (60 μm) microstructural analysis of a 130 m long ice core B40 drilled in Austral summer 2012/13 at the German Research station Kohnen on the East Antarctic plateau using artificial intelligence (AI)-enhanced micro-computed tomography (micro-CT) imaging and pore network modeling. Antarctic ice cores serve as valuable archives of past climate and environmental conditions. They consist of a sequence of compacted layers with varying density, ice and pore space structure, which are shaped differently depending on the climatic and environmental conditions during deposition and subsequent compaction. The in-depth development of microstructural features provides insights into the densification and pore closure processes as well as climate trends during the formation of the firn column.
However, traditional micro-CT imaging techniques applied to full core samples, although effective, often lack the necessary resolution to capture the intricate microstructure of ice samples fully. To address this limitation, we applied AI-driven super-resolution enhancement methods to improve the clarity and detail of micro-CT images, enabling a more precise quantification of ice core properties. Following AI enhancement, the study performed a comprehensive microstructure analysis on the full firn column, including geometrical, morphological, topological, and transport-related properties. The obtained metrics, including mean intercept length (MIL), cluster size, porosity (density), tortuosity, permeability, etc., can characterize the microstructural evolution of the ice column from snow to bubbly ice across different depths. The results reveal a systematic evolution in pore geometry, connectivity, and transport efficiency with depth, capturing the snow-to-ice transition with high fidelity. Fine and coarse layers were distinguished using the K-means clustering method, and anisotropy was detected in both the ice matrix and the pore space. Principal component analysis (PCA) showed that densification is influenced by multiple factors; in particular, during the pore close-off stage, variations in coordination number and throat dimensions led to distinct densification behaviors among the samples. These findings could contribute to the improvement of densification and air transport models by providing highly accurate data on the microstructure of ice cores and even enabling the definition of new microstructure-related climate proxy parameters.
Bagherzadeh et al. use microCT to conduct a microstructural analysis along a continuous 130 m-long ice core that captures the snow to ice transition in a region of the East Antarctica Plateau. Collecting three-dimensional and such large-volume data at high-resolution requires extensive scan time, which the study reduces by collecting low-resolution 3D data (in this case 120 um) and then applying a previously developed deep learning super-resolution method to increase 3D image resolution down to 60 um. In this study, the authors add to their method by applying additional steps in their data processing. The approach taken and the achievement of such a long continuous ice core analysis has potential to advance the documentation and characterization of ice core microstructures, as well as support interpretations of paleoclimate proxies contained within the ice. There is potential for contribution to our understanding of firn densification, pore close-off and gas-transport behavior.
General comments:
There is great potential for the authors’ approach and study to advance capabilities in reconstructing ice core records and for understanding the physical evolution of ice sheets as they transition from snow to ice. The manuscript will require much improvement primarily regarding writing structure, presentation (grammar, formatting), and graphical communication before it could be considered for publication. All grammatical and basic formatting errors should be corrected before a resubmission.
The explanation of the DLSR method could be improved, as many readers may not be well versed in AI methods at this point in time. It is right to reference your previous work, but readers would benefit from additional information that expresses the main ideas around your method. I suggest adding to Figure 1 a graphic that demonstrates physically/visually what is happening during the main steps in your method, particularly the “new” steps you present in this study. It can be difficult to imagine what is occurring in certain steps of the computational workflow. Can you put this into a graphical representation? (i.e. more than the flowchart with text).
As stated by the authors, a number of results from the microCT dataset are consistent with current views of the firn column and transition of snow to ice. However, they fail in many cases to reference the previous studies that have established those views. The advance in this study was not so obvious to me when reading the manuscript. Make clear what the new information gained is in the discussion/conclusion. Perhaps this is why it seems its impact is overstated but also consider whether you have overstated your claims.
There is terminology used that is misleading: “grain clusters” and “solid ice clusters” which are used synonymously. Please see a detailed comment below. If the authors are making a true inference about clusters of ice grains based on absorption contrast tomography (i.e. standard microCT, which cannot differentiate between grains), this would be a critical flaw in the analytical approach. I suggest clarifying your terminology so that it is consistent with what you are truly measuring. Otherwise, such issues propagate in the literature.
Specific comments (see additional specific comments and technical corrections in the marked-up PDF that is attached):