A DATA-DRIVEN APPROACH FOR SOIL PARAMETER DETERMINATION

USING SUPERVISED MACHINE LEARNING

Haris Felić
Institute of Soil Mechanics, Foundation Engineering and Computational Geotechnics, Graz University of Technology,
Austria. E-mail: [email protected]

Islam Marzouk
Institute of Soil Mechanics, Foundation Engineering and Computational Geotechnics, Graz University of Technology,
Austria. E-mail: [email protected]

Franz Tschuchnigg
Institute of Soil Mechanics, Foundation Engineering and Computational Geotechnics, Graz University of Technology,
Austria. E-mail: [email protected]

Soil constitutive models have significantly advanced over the years, often with an increase in parameters.
Accurate determination of these parameters is critical, as inaccuracies can lead to unreliable numerical
simulations. Conventional calibration practice typically relies on laboratory testing, which is often impractical
for applications, particularly at early project stages. An ongoing research project of the Computational
Geotechnics Group at Graz University of Technology focuses on developing an Automated Parameter
Determination framework that integrates a graph-based approach to derive constitutive model parameters from
in-situ tests. This framework uses established correlations to identify parameters based on in-situ tests. However,
the multiplicity of correlations for a given parameter introduces inherent challenge in selecting the recommended
value. An alternative method involves using advanced regression algorithms to enhance the robustness of
parameter determination through data-driven techniques. This approach can improve the quality of numerical
simulations and minimizes the uncertainty in the parameter calibration process. In this paper, supervised machine
learning regression models are employed to predict soil parameters, including saturated unit weight, undrained
shear strength, and small-strain shear modulus (via shear wave velocity) using cone penetration test data as input.
The performance of these models is benchmarked using data from two Norwegian GeoTest Site locations and
are validated based on laboratory-derived values, in-situ measurements, and traditional correlation-based
methods. The findings demonstrate the potential of advanced regression models to generalize soil parameter
predictions across test sites, significantly improving reliability while reducing site-specific biases.

Keywords: Constitutive soil parameter, machine learning, data-driven, site characterization, Cone penetration test

1. Introduction
Constitutive soil models approximate the non-linear behaviour of soils, which makes model selection and
parameter determination challenging. Generally, increasing model complexity raises the number of parameters.
One such parameter is the stress-dependent small-strain shear modulus G0, which depends on soil density ρt
and shear wave velocity vs. The latter can be obtained through in-situ (e.g. seismic cone penetration test (CPT))
or lab tests (e.g. bender elements). Other parameters, such as saturated unit weight γsat and undrained shear
strength su, are derived from lab tests (e.g. index tests for γ, triaxial tests for su) or in-situ tests (e.g. vane shear
tests for su). Ground conditions and basic soil properties in practical engineering are often initially assessed
using local geological knowledge and experience from similar sites. However, site investigations provide more
accurate, location-specific quantitative data. In geotechnical practice, in-situ data are often used in correlations
to estimate further soil parameters, requiring statistical analysis and engineering judgement (Phoon et al. 2022).
An ongoing research project focuses on developing an Automated Parameter Determination (APD) framework
using a graph-based approach, using established correlations to identify constitutive model parameters from
in-situ tests (Marzouk et al. 2024). Despite its potential, the multiplicity of correlations for a given parameter
makes it difficult to determine a single, recommended value. An alternative involves using advanced machine
learning (ML) regression algorithms to enhance the robustness of parameter determination through data-driven
approaches. This paper explores the application of ML models to improve the accuracy of predicting soil
parameters. The proposed framework uses CPT data to predict lab parameters such as γsat and su, as well as in-
situ measurements of vs. The study details the database, ML training process, ML model performance, and
validation using data from two Norwegian GeoTest Sites (NGTS) (L’Heureux and Lunne 2020).

2. Databases for Machine Learning

2.1. Databases for laboratory parameters
The database for laboratory parameters comprises pairs of CPT measurements and corresponding soil
parameters. These soil parameters are determined through lab tests conducted on soil samples extracted from

Proc. of the 9th International Symposium for Geotechnical Safety and Risk (ISGSR)
Copyright © 2025 by ISGSR 2025 Organizers. Published by Research Publishing, Singapore
ISBN: 981-973-0000-00-0 :: doi: 10.3850/981-973-0000-00-0- isgsr2025-paper
2 Felic, Marzouk, Tschuchnigg

boreholes located near the CPT tests. The database includes 384 pairs for γ and 84 pairs for su. Data sources
include publicly available data from Ballina data, NGTS (both accessible at geocalcs.com/datamap), and from
the Netherlands (see Lengkeek and Breedeveld 2022), and private data from Norway and Austria.
2.2. Databases for in-situ parameters
The in-situ database comprises 46 seismic CPTs and 254 seismic CPTu tests, resulting in 21,760 data points.
Data sources for this database include the New Zealand Geotechnical Database (Scott et al. 2015), Premstaller
Geotechnik (Oberhollenzer et al. 2021), Dutch site data (github.com/snakesonabrain/isc7_datasets), and
Taiwan site data (https://data.mendeley.com/datasets/v7frv3k2d3/1)

3. Machine Learning Models

The created ML models are discussed in this section, covering feature selection, the training workflow, and
model evaluation. This paper focuses exclusively on the use of one decision tree algorithm, namely XGBoost
(eXtreme Gradient Boosting Decision Tree), which is an enhanced version of the Gradient Boosting Decision
Tree (GBDT) algorithm. This algorithm focuses on improving computational speed and model accuracy by
introducing more regularization terms and parallelized tree building.
3.1. Feature selection
In ML, a model is a function that produces an output based on a given input (Deisenroth et al. 2021). Here, the
input matrix includes variables of depth, tip resistance qc, sleeve friction fs, friction ratio Rf, as also utilized in
Felić et al. (2024). Preliminary studies demonstrated that using depth, qc, fs, and Rf leads to good model
performance. In this framework, separate models are developed to predict each soil parameter: γsat, su, and vs.
3.2. Machine learning workflow training
The training workflow, following a process described by Deisenroth et al. (2021), begins with splitting the
dataset into an 80% training set and a 20% test set. The training set is exclusively used for model training,
while the test set is reserved for testing the model after model training. The initial training step starts using an
initial set of hyperparameters. After training, the model predictions are compared to the validation set (a subset
of the training data), and its performance (validation loss) is computed with an objective function. To minimize
overfitting, cross-validation is applied, where the data is split into K subsets, with K-1 used for training and
one for validation. The model's performance is averaged over these K iterations; in this study, ten-fold cross-
validation is employed. Additionally, early stopping is used to halt training if the validation error does not
improve for ten consecutive iterations. Hyperparameter optimization is performed using the Differential
Evolution algorithm from pymoo, with a population size of 25. This algorithm identifies the optimal
hyperparameters by maximizing the coefficient of determination R2 (see section 3.3) as the objective function.
The optimization process runs for up to 1000 iterations or until a maximum relative error of 10−4 is achieved.
Once the hyperparameters are optimized, the model is retrained using the combined training and validation
data. The model performance is then evaluated on the 20 % test set. If the test loss is satisfactory, a final training
phase is conducted with the entire dataset using the best hyperparameters obtained from the optimization.
3.3. Evaluation of model performance
For evaluation and optimization, the coefficient of determination R2 serves as the objective function. R2 is
expressed as:
𝑅𝑅2 = 1 − ∑(𝑦𝑦𝑖𝑖 − 𝑓𝑓𝑖𝑖 )2 / ∑(𝑦𝑦𝑖𝑖 − �𝑦𝑦)2 (1)
Here, yi denotes the true values, fi represents the predicted values, and ȳ is the mean of the true values. R2
measures the predictive accuracy of a model, typically ranging from 0 to 1. A higher R2 value indicates better
alignment between the model's predictions and the observed data, while a lower R2 suggests poorer alignment.

Fig. 1. ML model performance after hyperparameter optimization with ten-fold cross validation and early stopping.
 Proc. of the 9th International Symposium for Geotechnical Safety and Risk (ISGSR) 3

Fig. 1 illustrates the performance of the ML model after hyperparameter optimization, focusing on the test loss
evaluation. Each soil parameter (γsat, su, vs) in Fig. 1 shows the predicted data plotted against the actual data (as
grey scatter points). The dashed 45° line represents perfect alignment between the actual and predicted values.
The scatter distributions indicate strong overall performance for most predictions, with some instances of
overestimation and underestimation. For example, some vs predictions between 0-350 m/s are either under- or
overestimated compared to the actual data. The R2 values for the three soil parameters range from 0.58 to 0.89.

4. Application of Machine Learning Models

This section presents the performance of the trained ML model using real CPT data from NGTS (L’Heureux
and Lunne 2020). These sites encompass various soil types, including clay, silt, quick clay, sand, and
permafrost. For this study, only the silt and clay sites are considered. The ML model's performance is evaluated
by comparing its predictions with laboratory data (index, triaxial, and direct shear tests), in-situ tests (seismic
dilatometer tests, SCPTu, CPTu) from the selected test sites, and the following established correlations:
C1 Robertson and Cabal (2010) γ𝑠𝑠𝑠𝑠𝑠𝑠 = γw (0.27(log R f ) + 0.36(log qt /pa ) + 1.236) (2)
C2 Lengkeek and Brinkgreve (2022) γ𝑠𝑠𝑠𝑠𝑠𝑠 = 19.5 − 2.87 ∗ (log(9000/qt ) /log (20/R f )) (3)
C3 Mayne (2016) su = (qt − u2 )/Nke ; Nke = 8 (4)
C4 Mayne (2016) su = (qt − σv )/Nkt ; Nkt = 12 (5)
C5 Mayne (2006) vs = 51.6 ∗ ln(fs ) + 18.50 (6)
C6 Hegazy and Mayne (1995) vs = 12.02 ∗ q0.319
c ∗ fs−0.0466 (7)
4.1 Norwegian GeoTest Sites – Silt site
Fig. 2 presents the CPTu data from the silt site in Halden, alongside ML predictions, lab and in-situ site data,
and correlations for each soil parameter: γsat, su, vs, and G0. G0 was calculated as vs2 ∙ γsat/9.81 for both
correlations and ML predictions. The grey hatched regions in each figure indicate the upper and lower bounds
of the selected correlations. For γsat, vs, and G0, these regions reveal scatter, and a significant offset compared
to the lab and in-situ data. su also exhibits a larger range and offset relative to the lab data. For su, the total
stress 𝜎𝜎𝑣𝑣 using Eq. (5) was calculated based on the average γsat from Eq. (2) and (3). The ML predictions for
γsat, vs, and G0 show a very good agreement with the lab and in-situ data across the entire depth profile. However,
predictions for su tend to overestimate the lab values, although the increasing trend of su with depth is well
captured compared to correlations and lab data. The reason for this overestimation is part of ongoing research.
Comparing both methods (correlations, and ML predictions), ML models tend to be less site-specific as the
training database contains no data from NGTS, whereas correlations inherently incorporate site-specific
characteristics, making it often difficult to determine which correlation is most suitable for a given site.

Fig. 2. Raw CPTu site data for HALC11 (dashed black line), along with the ML prediction (solid black line), lab and in-
situ data (represented by circular and triangular markers, respectively), and correlations (dotted black line).
4.2 Norwegian GeoTest Sites – Clay site
Fig. 3 illustrates SCPTu data from the clay site in Onsøy, ML predictions, lab and in-situ site data, and
correlations for each soil parameter: γsat, su, vs, and G0. G0 was calculated similar to the silt site. The grey hatched
regions represent the bounds of the selected correlations. γsat and su correlations align well with the lab data,
showing minimal scatter. For vs, and G0, minor scatter and slight offsets are observed compared to the in-situ
data. ML predictions for γsat closely match lab data along depth. Predictions for vs and G0 align well with the
in-situ data between 1–7 m, although an overestimated trend occurs below this depth. The overestimation is
most likely due to sparse datapoints (CPT data) within the database for this material. su predictions also align
well with the lab data, but exhibit a sensitive, zig-zag pattern with increasing depth. Overall, the ML models
appear to be less sensitive to site-specific conditions, comparable to the findings at the silt site, whereas
correlations may be influenced by site-specific characteristics.
4 Felic, Marzouk, Tschuchnigg

Fig. 3. Raw SCPTu site data for ONSC08 (dashed black line), along with the ML prediction (solid black line), lab and in-
situ data (represented by circular and triangular markers, respectively), and correlations (dotted black line).

5. Conclusion
This study demonstrates the potential of supervised machine learning (ML) models to enhance soil parameter
determination from cone penetration test data, addressing the limitations in traditional empirical correlations.
By employing advanced regression algorithms, such as XGBoost, data-driven approaches significantly
improve prediction accuracy for soil parameters, including saturated unit weight, undrained shear strength, and
small-strain shear modulus (derived from shear wave velocity). Validation using data from the Norwegian
GeoTest Sites highlights the robustness of these ML models in capturing soil parameters across different
locations. While slight overestimations occur at certain depths, the models provide generalizable predictions
and can effectively reduce uncertainty in parameter calibration. These advancements are particularly valuable
in scenarios requiring robust and reliable soil characterization, especially where limited data is available. The
proposed framework will be incorporated into the Automated Parameter Determination research project as an
additional information source to enhance parameter accuracy. Future work will focus on database extensions,
managing outliers, and investigating potential biases that may affect the ML performance. The open-source
availability of both datasets and algorithms on GitHub ensures transparency and fosters further collaborations.

Acknowledgement
The authors express their gratitude to Simon Oberhollenzer for providing a subset of the laboratory database.

Data repository
The GitHub repository is available here: https://github.com/harifel/ISGSR25_DataDrivenSiteCharacterization/

References
Deisenroth, M. P., Ong, C. S., and Faisal, A. A. (2021). Mathematics for machine learning. Cambridge University Press.
Felić, H., Marzouk, I., Peterstorfer, T., Tschuchnigg, F. (2024). Data-driven site characterization - Focus on small-strain
stiffness. In M. Arroyo, A. Gens, 7th ISC – Ground models, from big data to engineering judement.
Hegazy, Y. A. and Mayne, P. W. (1995). Statistical correlations between Vs and cone penetration data for different soil
types. In Swedisch Geotechnical Society (Eds.), International Symposium on Cone Penetration Testing, CPT‘ 95.
L’Heureux, J.-S., Lunne, T. (2020). Characterization and Engineering properties of Natural Soils used for Geotesting, AIMS
Geosciences Volume (6). https://doi.org/10.3934/geosci.2020004.
Lengkeek, H. J., Breedeveld, J. (2022). Eemdijk full-scale test on dike reinforced by sheet pile. Dataset
https://doi.org/10.4121/19213890.v1.
Lengkeek, H. J. and Brinkgreve, R. (2022). CPT-based unit weight estimation extended to soft organic clays and peat: An
update. In L. Tonni, G. Gottardi (Eds.), Cone Penetration Testing 2022. https://doi.org/10.1201/9781003308829-71.
Marzouk, I., Brinkgreve, R., Lengkeek, A., and Tschuchnigg, F. (2024). APD: An automated parameter determination
system based on in-situ tests. Computers and Geotechnics 176. https://doi.org/10.1016/j.compgeo.2024.106799.
Mayne, P. (2014). Interpretation of geotechnical parameters from seismic piezocone tests. In Robertson, P. K., Cabal, K.
L. (Eds.), Proceedings from the 3rd International Symposium on Cone Penetration Testing. CPT‘ 14.
Mayne, P. (2016). Evaluating effective stress parameters and undrained shear strengths of soft-firm clays from CPTu and
DMT. In Lehane B. M., Acosta-Martinez H. E., Kelly R. (Eds.), Geotechnical and Geophysical Site Characterization 5.
Oberhollenzer, S., Premstaller, M., Marte, R., Tschuchnigg, F., Erharter, G. H., and Marcher, T. (2021). Cone penetration
test dataset Premstaller Geotechnik, Data in Brief 34. https://doi.org/10.1016/j.dib.2020.106618.
Phoon, K.-K., Ching, J., and Shuku, T. (2022). Challenges in data-driven site characterization. Georisk: Assessment and
Management of Risk for Engineered Systems and Geohazards 16. https://doi.org/10.1080/17499518.2021.1896005.
Robertson, P. K., & Cabal, K. L. (2010). Estimating soil unit weight from CPT 2010. In Robertson, P. K., P.W. Mayne
(Eds.), Proceedings of the 2nd International Symposium on Cone Penetration Testing.
Scott, J. W., van Ballegooy, S., Stannard, M., Lacrosse, V., et al. (2015). The Benefits and Opportunities of a Shared
Geotechnical Database. In Cubrinovski (Ed.), 6th International Conference on Earthquake Geotechnical Engineering.