The research, led by Zixuan Yu and colleagues working across the China University of Mining and Technology, the University of Nottingham and Nantong University, tackles a trio of long-standing problems in weathering assessment. Conventional methods typically demand destructive laboratory testing, rely on expert analysts interpreting microscopy images by eye, and struggle to capture damage in its earliest stages, when intervention is still possible and inexpensive. To overcome these limitations, the team developed a deep learning framework capable of automatically identifying the deterioration level of salt-weathered granite directly from electron micrographs, removing much of the human subjectivity that has historically clouded such evaluations.

The experimental foundation of the work was as demanding as it was methodical. Granite specimens were subjected to repeated wet-dry cycling in sodium sulfate solution, a procedure designed to mimic the harsh conditions experienced by rock in salt-rich environments where fluctuating moisture drives repeated cycles of salt dissolution and crystallization. These crystallization pressures pry open pores, detach clay minerals and propagate microcracks, progressively degrading the rock’s load-bearing skeleton. From the cycled specimens, the researchers collected more than 5,000 scanning electron microscope images spanning five distinct deterioration levels, creating a large labeled dataset that traces the visual evolution of weathering from pristine crystalline texture to heavily degraded, crack-riddled fabric.

Training a neural network on this corpus produced striking results. The deep learning model achieved a maximum classification accuracy of 89.1 percent when analyzing images at an observation scale of 12.7 by 12.7 micrometers, a field of view fine enough to resolve individual mineral grains and their boundary conditions. Interestingly, the model proved more reliable at distinguishing the extremes of the deterioration spectrum than the middle. Intact and severely weathered specimens, with their sharply contrasting microstructural signatures, were classified with higher distinctiveness than intermediate stages, whose transitional microstructures blur the visual boundaries between adjacent grades. This behavior mirrors the very ambiguity that challenges human experts, but crucially the framework offers a systematic route to managing it.

Recognizing that real-world predictions rarely fall cleanly into discrete categories, the team introduced a probability correction method grounded in historical data. When the classifier reports, for instance, that a specimen is most likely at deterioration level three but with substantial probability mass assigned to levels two and four, the correction scheme adjusts the final estimate to compensate for systematic overestimation and underestimation observed in past predictions. This statistical refinement adds a layer of reliability to the framework, acknowledging that weathering is a continuous process being squeezed into categorical labels, and that honest uncertainty quantification is essential for engineering decision-making.

Perhaps the most compelling aspect of the study is its insistence on interpretability. Deep learning models are often criticized as black boxes, particularly in fields like geotechnical engineering where stakes are high and physical plausibility matters. To open the box, the researchers applied deep Taylor decomposition, an explainability technique that attributes the network’s classification decisions to specific regions of the input image. The analysis revealed that the model was not latching onto irrelevant artifacts but was instead concentrating its attention on physically meaningful features of weathering damage: the detachment of clay minerals from the rock fabric, the exposure of grain boundaries, and the expansion of micropores and microcracks. In other words, the artificial intelligence had learned to see the damage in much the same way a trained petrographer would.

To provide external validation of this interpretation, the team extracted feature maps from the network and computed their fractal dimensions, comparing these values with the surface roughness of the weathered specimens. The results were persuasive. The fractal dimensions of the model-extracted feature maps correlated more strongly with surface roughness than those derived from the original SEM images, achieving a coefficient of determination of 0.81. Surface roughness is a well-established macroscopic indicator of weathering intensity, so this tight statistical relationship supplies independent physical evidence that the features the network learned are genuinely tied to the deterioration process rather than being statistical curiosities. It is a rare and valuable bridge between machine-learned representations and measurable material properties.

The implications extend well beyond the laboratory. Tunnels, dam foundations, retaining structures and heritage stone facades in coastal, arid and saline environments all face the relentless chemistry of salt attack, and deterioration that goes undetected at the microscopic level eventually manifests as costly failures. The study’s authors note that recent engineering literature documents collapses in weathered granite fault zones and accelerated tool wear in mixed weathered strata, underscoring the practical cost of underestimating rock degradation. A microstructure-informed evaluation tool that requires only electron micrographs, rather than destructive mechanical testing, could support earlier diagnosis, more targeted maintenance schedules and more reliable durability assessments for infrastructure and conservation projects alike.

The work also situates itself within a rapidly growing research movement that brings computer vision to materials science. Deep learning approaches have already demonstrated success classifying steel microstructures, identifying rock lithology from microscopic images and detecting cracks in civil structures, and the present team has previously applied similar frameworks to rocks degraded by high temperatures. By extending these techniques to salt weathering, one of the most widespread and least quantified deterioration mechanisms in geomaterials, the study fills an important gap. The authors caution that their model shows reduced distinctiveness for intermediate weathering stages, suggesting that future refinements, larger datasets and possibly multi-scale imaging could further improve performance across the full deterioration spectrum.

Still, the trajectory is clear. What once required a seasoned expert hunched over a microscope, judging weathering grades by intuition and experience, can now be performed rapidly and reproducibly by an algorithm that explains its own reasoning. With accuracy approaching 90 percent, physically interpretable feature attribution, and a validated statistical link to surface roughness, the framework demonstrates that intelligent microscopy can quantify, objectify and illuminate the earliest fingerprints of salt weathering in granite. For engineers tasked with protecting structures in unforgiving environments, and for conservators racing to preserve weathered stone heritage, that combination of speed, transparency and early-stage sensitivity may prove transformative.

Subject of Research: Deep learning-based identification of salt weathering deterioration in granite using SEM micrograph classification

Article Title: Intelligent evaluation of rock weathering deterioration using sem characterization and deep learning recognition

Article References: Yu, Z., Jing, H., Su, X., Lu, X., Miao, J., & Gao, Y. (2026). Intelligent evaluation of rock weathering deterioration using sem characterization and deep learning recognition. Earth Science Informatics, 19(11), Article 191. https://doi.org/10.1007/s12145-026-02242-3

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02242-3

Keywords: deep learning, rock weathering, salt weathering, granite, SEM image classification, sodium sulfate, microstructure, fractal dimension, surface roughness, geotechnical engineering, deep Taylor decomposition, deterioration assessment