A data-driven review shows how modern machine-learning models consistently outperform traditional empirical equations in predicting soil shear modulus and damping ratio, offering geotechnical engineers a clearer, more efficient path to characterizing dynamic soil behavior while highlighting practical limitations, data requirements, and future research needs.
Why Soil Dynamics Still Challenge Engineers
Accurate characterization of soil dynamic properties remains one of the more demanding tasks in geotechnical engineering. Two parameters are central to this challenge: shear modulus (G), representing soil stiffness, and damping ratio (D), describing the soil’s ability to dissipate energy.
These properties govern soil response under dynamic loading conditions such as earthquakes, machine vibrations, and traffic. They directly influence site response analyses, foundation performance, slope stability, and soil–structure interaction. Inadequate estimates of G and D can therefore introduce uncertainty into design, with implications for both safety and cost.
Conventional determination of these parameters relies on laboratory testing—such as cyclic triaxial, cyclic simple shear, and resonant column tests—or in-situ methods including downhole, crosshole, and seismic cone penetration testing. While well established, these methods face recognized limitations. Laboratory tests may suffer from sample disturbance, scale effects, and difficulties in reproducing in-situ stress conditions, while field tests can be constrained by access, equipment, and environmental variability. Both approaches are resource-intensive, particularly when multiple tests are required.
Empirical correlations derived from experimental data have long been used as practical alternatives, but their ability to capture the nonlinear and highly variable nature of soil behavior is limited. Reported prediction accuracies for these equations often fall below R² = 0.85, especially when applied beyond their original calibration ranges.
What This Review Examines
Rather than proposing a new predictive model, the reviewed study takes a broader perspective. It presents a systematic review and quantitative synthesis of machine-learning (ML) approaches used to predict soil shear modulus and damping ratio.
Unlike earlier reviews that considered ML in geotechnical engineering in general terms or focused on individual algorithms, this work compiles results across soil types, testing methods, and modeling families. The authors compare ML-based predictions directly with established empirical equations and examine how model performance varies with dataset size and soil characteristics.
The review draws on more than 50 ISI-indexed journal papers, supplemented where necessary by conference studies. Only studies reporting sufficient methodological detail—including dataset characteristics, validation procedures, and performance metrics—were included, allowing meaningful comparison across approaches.
Machine Learning Methods in Context
The review covers a wide range of ML techniques used in soil dynamics, including Artificial Neural Networks (ANN), Support Vector Machines (SVM and SVR), tree-based ensemble models such as Random Forest and XGBoost, Evolutionary Algorithms, and Fuzzy Logic-based systems.
Across these families, model performance was assessed using commonly reported metrics such as R², RMSE, MAE, and MSE. The authors also examined reported advantages—such as the ability to capture nonlinear interactions—as well as limitations, including sensitivity to dataset size and data quality.
Importantly, the review emphasizes that predictive success depends not only on the choice of algorithm, but also on data preprocessing, feature selection, and validation strategy. Stress-related variables, particularly confining pressure, and state parameters such as void ratio were consistently identified as influential inputs across soil types.
What the Evidence Shows
Across the surveyed literature, machine-learning models consistently achieved higher prediction accuracy than traditional empirical equations for both shear modulus and damping ratio.
Reported R² values for ML models reached as high as 0.994, compared with empirical correlations that typically remained below 0.85. Artificial Neural Networks were the most frequently applied models and performed well across a wide range of soils. Support Vector Machines demonstrated strong generalization capabilities, particularly for moderate-sized datasets. Tree-based ensemble models showed robust performance and were effective across both small laboratory datasets and larger compiled databases.
The review does not suggest that a single ML method is universally superior. Instead, it highlights that model performance depends on soil type, dataset size, and the specific dynamic property being predicted. Nevertheless, the collective evidence indicates that ML approaches offer a meaningful improvement in predictive capability when compared with conventional empirical methods.
Practical Implications and Future Directions
From a practical standpoint, the findings suggest that machine-learning models can serve as efficient predictive tools, potentially reducing the need for extensive laboratory testing when suitable datasets are available. However, the authors also emphasize that ML models should be applied with care, particularly in design contexts where uncertainty must be explicitly managed.
The review identifies several areas requiring further development, including the expansion of geographically diverse databases, improved field validation, and the integration of physics-informed features to enhance model generalizability. These steps are viewed as essential before ML-based predictions can be more widely adopted in routine geotechnical design practice.
Acknowledgment and Reference
We acknowledge the contribution of Samira Ghorbanzadeh, Danial Jahed Armaghani, and their co-authors for providing a comprehensive and balanced assessment of machine-learning applications in soil dynamics.
Reference
Ghorbanzadeh, S., Jahed Armaghani, D., Salimi, M., Payan, M. (2026). Predictive models for dynamic properties of soils using machine learning approaches: A comprehensive review. Engineering Applications of Artificial Intelligence, 163, 113014.