Published in the journal AI in Civil Engineering, the study introduces a hybrid framework that combines acoustic diagnostics with machine learning to automate the assessment of frost heave risks. The system aims to reduce the reliance on traditional, labor-intensive geotechnical methods such as borehole sampling and extensive laboratory testing, which often fail to capture the spatial variability of soil properties in linear infrastructure projects.
The central objective of the study was to create a robust AI-agent that could characterize the relative acoustic compressibility coefficient, denoted as β, based on the physical and mechanical properties of soil. This coefficient serves as a key indicator of soil structure and its susceptibility to deformation under freeze-thaw cycles. By integrating spectral reflectance data with machine learning algorithms, the team sought to provide a transformative approach to soil assessment that offers real-time, large-scale characterization without the need for invasive sampling.
The experimental setup employed digital techniques to process and record longitudinal and transverse acoustic pulse signals reflected from piezoelectric sensors mounted on soil specimens. These signals were automatically transferred via a USB adapter to a personal computer for further analysis by the AI-agent. Acoustic diagnostics were performed using Fast-Fourier Transform (FFT) Spectral Analysis, which allowed the researchers to correlate waveform spectra with heaving deformation. The temperature regimes tested ranged from 0°C to -5°C and back, simulating the natural freeze-thaw dynamics observed in temperate continental climates.
The AI-agent utilized a hybrid architecture combining a Convolutional Neural Network (CNN), a Support Vector Machine (SVM), and a Random Forest (RF). The CNN was responsible for extracting high-level feature maps from the spectral data, capturing the physical changes in soil structure. The SVM was employed to classify different soil types based on their spectral signatures and moisture content, while the RF model captured non-linear relationships among soil parameters to predict heaving potential. This ensemble approach was designed to address the complexities of heterogeneous soil data and mitigate overfitting, a common challenge in geotechnical machine learning applications.
Soil specimens were collected from two distinct geoclimatic zones along the prospective Moscow–Kazan rail corridor. Site 1, located in the south-eastern Moscow region, featured Holocene alluvial deposits with low to moderate frost susceptibility. Site 2, situated in the Republic of Tatarstan near Kazan, consisted of lacustrine-glacial formations with more pronounced frost heave deformations. A total of 150 soil specimens were extracted from each site for non-destructive spectral analysis and laboratory soil heaving testing. The laboratory tests followed GOST 28622-2012 standards, quantifying soil deformation resulting from frost heaving under controlled freezing and thawing cycles.
To expand the dataset for AI agent training, the researchers employed the Random Forest technique to generate an additional 700 synthetic samples, resulting in a total of 1,000 data samples for analysis. This data augmentation strategy was bounded by domain-specific physical thresholds, such as moisture content limited to 12%–26% and clay content less than 45% by weight. The synthetic data were validated by comparing their statistical distributions with those of the original experimental data, ensuring that the generated samples closely aligned with the physical characteristics of the real soil datasets.
The results indicated that as the density and skeletal density of the soil increased, the velocities of elastic waves rose, while the ratio of longitudinal to transverse wave velocities decreased. Acoustic compressibility (β) increased linearly with frequency in sand and sandy loam, whereas loam and clay exhibited more complex, progressive increases due to their higher energy absorption. The AI-agent demonstrated promising performance, with its soil heaving potential classifications aligning closely with those derived from laboratory and on-site methods. However, minor discrepancies were observed for loam and clay, potentially attributed to equipment sensitivity and the presence of outlier samples in the dataset.
In terms of classification accuracy, the macro-averaged precision reached 0.89, recall 0.87, and F1-score 0.88. The model showed particularly strong performance in detecting “Moderate” and “High” heaving levels, with F1-scores exceeding 0.91. When aggregated into binary outcomes of “Failure” and “Safe,” the AI-agent achieved a precision of 0.92 and a recall of 0.91 for “Failure” detection. This reliability in identifying high-risk frost-heaving zones is critical for the safety and resilience of rail transport infrastructure, enabling early detection of potential structural damage.
The study also highlighted the economic advantages of the AI-agent framework. The system was deployed locally on a PC with a total hardware cost of approximately $4,400, including acoustic sensors and diagnostic hardware. In comparison, seasonal field campaigns typically cost over $14,000–$20,000 per season, and laboratory campaigns approach $12,000–$16,000. The one-time operational cost for the AI-agent amounted to $1,000–$1,500, covering the compensation for a specialist serving as a prompt engineer and coordinator. This cost-effectiveness, combined with real-time inference capability of approximately 3.4 seconds per instance, makes the system a viable alternative to traditional assessment methods.
Despite its high performance, the framework has several limitations. The training data are inherently biased toward soil types dominant in the Moscow–Kazan region, which may limit direct transferability to drastically different pedological contexts such as arid calcareous soils or tropical laterites. Additionally, the agent was trained on a thermal envelope of −5 °C to 0 °C, which may not account for the deeper permafrost or higher freeze-thaw amplitudes found in other climatic zones. The study acknowledges that further refinement of the training dataset and the incorporation of additional physical features, such as temperature gradients and soil microstructure, could enhance the model’s robustness and generalization capability.
To ground the acoustic measurements in physical principles, the investigation adapted a single-degree-of-freedom harmonic oscillator model that linked wave propagation to the elastic and inertial macro-properties of the soil skeleton. Under the assumption that the medium behaves homogeneously within the 100 Hz to 20 kHz operating range, the team calculated natural oscillation frequencies based on source stiffness and attached mineral mass. For a spherical mineral emitter with radius R and density ρ, this yielded the acoustic compressibility coefficient, formulated as β = 4π²ρ(f₀²R) in units of newtons per cubic meter. Experimental ultrasonic impulse testing between 20 kHz and 120 kHz on twelve-centimeter soil monoliths allowed researchers to capture longitudinal and transverse wave delay times using a digital oscilloscope. These measurements supplied direct values for the primary wave velocity and transverse wave velocity, enabling the team to evaluate dynamic elastic constants, specifically Young’s modulus and shear modulus, across controlled thermal shifts.
Fast-Fourier Transform analysis revealed clear acoustic windows and damping characteristics unique to each soil texture. Sandy soils demonstrated broader spectral peaks because low cohesion fostered greater energy dissipation, while mixed wavefront signals separated into three distinct resonance points. The lower resonance at 25 kHz captured bulk elastic properties, where weak amplitudes reflected low bulk density and elevated compressibility. The intermediate peak at 42 kHz corresponded to pore spaces and particle packing density, whereas the 72 kHz resonance reflected moisture levels and the stiffness of individual grain contacts. When predicting structural instability, the model determined critical threshold limits for acoustic compressibility. Loam registered a threshold value of 11.6 × 10⁹ N/m³ across frequencies of 47 kHz to 90 kHz at a particle density-radius product of 0.15 kg/m², while clay reached 10.5 × 10⁹ N/m³ between 52 kHz and 80 kHz at 0.65 kg/m², establishing loam as the most structurally unstable soil examined.
Field validation relied on an extensive monitoring network established across the two prospective rail corridor sites over the course of one full year. Field crews drilled five boreholes at depths of 0.20, 0.40, 0.60, 0.80, and 1.00 meter per site to house heave meters, alongside a five-meter borehole cased in polyvinyl chloride to track subsurface temperatures and reference benchmarks anchored at three meters. The minimum recorded field temperatures dropped to -1.3 °C at Site 1 and -0.8 °C at Site 2, demonstrating that the laboratory envelope of -5 °C provided a rigorous margin of safety for severe seasonal freezing. Correlation matrices across thermal regimes confirmed distinct phase behaviors: during freezing from 0 °C to -5 °C, sand and sandy loam showed a strong correlation of 0.85, whereas loam and clay exhibited a correlation of 0.70. During thawing back toward 0 °C, correlations rose to 0.86 for coarse soils and 0.75 for fine-grained soils, illustrating how moisture retention and ice crystal formation govern acoustic compressibility during phase transitions.
To support future scaling across varied transportation networks, the system architecture incorporates an Agent-to-Agent protocol facilitating modular inter-agent communication through structured JSON messages over local sockets or message-passing interfaces. This protocol enables the primary soil heaving prediction agent to coordinate with planned supplementary modules, including an environmental monitor agent, a spectral integrity agent, an acoustic interpreter, and an infrastructure response agent. These specialized modules are designed to recalibrate feature extractors when encountering acoustic drift caused by organic matter or salinity, while containerized deployment allows edge-to-cloud execution.
Subject of Research: AI-driven spectral analysis of soil heaving for automated surveys in rail transport infrastructure
Article Title: AI-driven spectral analysis of soil heaving for automated surveys in rail transport infrastructure
Article References: Zaitsev, A., Koshurnikov, A., Gagarin, V., Frolov, D., & Rzhanitsyn, G. (2025). AI-driven spectral analysis of soil heaving for automated surveys in rail transport infrastructure. AI in Civil Engineering, 4(1), Article 23. https://doi.org/10.1007/s43503-025-00072-8
Image Credits: AI Generated
DOI: 10.1007/s43503-025-00072-8
Keywords: AI-driven, spectral, analysis, soil, heaving, automated, surveys, rail, transport, infrastructure, scientific research
News Source: Denise Maddox. (November 10, 2026). AI Agent Predicts Soil Heaving for Rail Infrastructure. Scienmag.
