The mechanical characterization of soils and the prediction of their material parameters through experimental procedures are important aspects of civil engineering. Nonetheless, these procedures can be costly and difficult to perform, whereas the neural network framework offers a computational tool for predicting material parameters from small datasets. In this paper, predictions of elasticity parameters such as Young’s modulus, shear strength parameters including the friction angle and cohesion, and hydraulic flow parameters such as the consolidation coefficient <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$C_{v}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>C</mi> <mi>v</mi> </msub> </math></EquationSource> </InlineEquation> are made using feed-forward neural networks trained on in situ data collected in Heraklion, Crete, Greece. The input parameters include the void ratio e, percentage of the humidity of the soil w, and the specific weight of the soil skeleton <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\gamma _{s}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>γ</mi> <mi>s</mi> </msub> </math></EquationSource> </InlineEquation>. The analysis of model hyperparameter estimation and convergence during the supervised learning process indicates that approximately 10 epochs are sufficient, with the relative root mean square error (RRMSE) remaining below <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$10^{-7}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mn>10</mn> <mrow> <mo>-</mo> <mn>7</mn> </mrow> </msup> </math></EquationSource> </InlineEquation> for all models except the one predicting unconfined compressive strength, where the RRMSE is below <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$10^{-2}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mn>10</mn> <mrow> <mo>-</mo> <mn>2</mn> </mrow> </msup> </math></EquationSource> </InlineEquation>. Furthermore, the results indicate that when <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$\gamma _{s}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>γ</mi> <mi>s</mi> </msub> </math></EquationSource> </InlineEquation> increases, the behavior of most output variables is not strictly monotonic, with subsets of both increasing and decreasing trends. The maximum deviation observed may reach approximately 150%, based on the ratio of maximum to minimum predicted values. Within the context of the available site-specific dataset, the proposed models demonstrate interpretability and consistent predictive performance, while remaining adaptable to future data updates. However, they remain limited to the specific site under study, and generalization to broader areas would require larger datasets, including additional experimentally obtained samples beyond the synthetically augmented dataset of the present study. The proposed model, due to the limited dataset size, should be regarded as a practical proof-of-concept framework rather than a fully generalizable computational tool. Finally, the model predictions were validated through comparison with values reported in the literature for cohesive soils of medium stiffness and strength.

A Prediction of Cohesive Soil Material Parameters, Situated in Heraklion, Crete-Greece, with the Implementation of Machine Learning Methods

期刊论文
出版模式：

开放获取
发表日期：

2026年5月29日

查看全文

Ambrosios-Antonios Savvides,
Andreas Antoniou

摘要

The mechanical characterization of soils and the prediction of their material parameters through experimental procedures are important aspects of civil engineering. Nonetheless, these procedures can be costly and difficult to perform, whereas the neural network framework offers a computational tool for predicting material parameters from small datasets. In this paper, predictions of elasticity parameters such as Young’s modulus, shear strength parameters including the friction angle and cohesion, and hydraulic flow parameters such as the consolidation coefficient $C_{v}$ $C_{v}$ are made using feed-forward neural networks trained on in situ data collected in Heraklion, Crete, Greece. The input parameters include the void ratio e, percentage of the humidity of the soil w, and the specific weight of the soil skeleton $\gamma _{s}$ $γ_{s}$ . The analysis of model hyperparameter estimation and convergence during the supervised learning process indicates that approximately 10 epochs are sufficient, with the relative root mean square error (RRMSE) remaining below $10^{-7}$ $10^{- 7}$ for all models except the one predicting unconfined compressive strength, where the RRMSE is below $10^{-2}$ $10^{- 2}$ . Furthermore, the results indicate that when $\gamma _{s}$ $γ_{s}$ increases, the behavior of most output variables is not strictly monotonic, with subsets of both increasing and decreasing trends. The maximum deviation observed may reach approximately 150%, based on the ratio of maximum to minimum predicted values. Within the context of the available site-specific dataset, the proposed models demonstrate interpretability and consistent predictive performance, while remaining adaptable to future data updates. However, they remain limited to the specific site under study, and generalization to broader areas would require larger datasets, including additional experimentally obtained samples beyond the synthetically augmented dataset of the present study. The proposed model, due to the limited dataset size, should be regarded as a practical proof-of-concept framework rather than a fully generalizable computational tool. Finally, the model predictions were validated through comparison with values reported in the literature for cohesive soils of medium stiffness and strength.