See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/379219532

A Case Study on Advanced CPT Data Interpretation: From Stratification to Soil

Parameters

Article in Geotechnical and Geological Engineering · March 2024

DOI: 10.1007/s10706-024-02774-9

CITATIONS READS

3 483

4 authors:

Islam Marzouk Andreas-Nizar Granitzer

Graz University of Technology Graz University of Technology
14 PUBLICATIONS 11 CITATIONS 17 PUBLICATIONS 48 CITATIONS

SEE PROFILE SEE PROFILE

Stefan Rauter Franz Tschuchnigg

Graz University of Technology Graz University of Technology
4 PUBLICATIONS 46 CITATIONS 235 PUBLICATIONS 894 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Andreas-Nizar Granitzer on 23 March 2024.

The user has requested enhancement of the downloaded file.

Geotech Geol Eng
https://doi.org/10.1007/s10706-024-02774-9

ORIGINAL PAPER

A Case Study on Advanced CPT Data Interpretation: From

Stratification to Soil Parameters
Islam Marzouk · Andreas‑Nizar Granitzer ·
Stefan Rauter · Franz Tschuchnigg

Received: 5 April 2023 / Accepted: 19 February 2024

© The Author(s) 2024

Abstract The cone penetration test (CPT) is consid- determine constitutive model parameters from in-situ
ered as one of the most reliable in-situ tests and has tests using a graph-based methodology. In the present
found numerous applications in the geotechnical engi- work, the developed automated parameter determina-
neering field. Traditional CPT interpretation includes, tion framework is applied to evaluate the soil param-
but are not limited to the identification of the soil eters of one selected soil layer identified from the
stratification and the determination of soil parameters. CPT interpretations. Potential lines of research in the
This paper presents a case study concerning a test site context of CPT interpretation are explored through-
located in Salzburg, Austria, in which we focus on the out this work and may serve as valuable reference in
interpretation of CPTs from different perspectives. future research.
The manuscript is divided into three main sections
dealing with three different aspects of CPT interpreta- Keywords CPT · Machine learning · Parameter
tion, namely stratification, ground variability and soil determination · Stratification
parameters. The first strategy introduces a machine
learning based stratification identification strategy to
detect soil layer boundaries from CPT measurements. 1 Introduction
A comparison with reference solutions demonstrates
relative merits of this approach to classical filter Since its introduction to the geotechnical engineer-
algorithms based on empirical CPT classifications. ing field the cone penetration test (CPT) has been
The second strategy introduces an intuitive approach applied to numerous engineering tasks, such as
to evaluate the ground variability. This is achieved the estimation of bearing capacity, design of shal-
by calculating the level of fluctuation on the basis low and deep foundations, walls and embankments,
of CPT measurements and could be used as a data- as well as the assessment of ground improvement
driven decision-making tool for the improved design measures and the site resistance to liquefaction
of CPT investigation layouts. The third strategy is (Umar et al. 2018; Kumar et al. 2021). Neverthe-
embedded in an ongoing research project that aims to less, estimating soil parameters and site charac-
terization may be regarded as the main purposes
I. Marzouk (*) · A.-N. Granitzer · S. Rauter · for carrying out CPTs (Lunne et al. 1997; Schnaid
F. Tschuchnigg 2009; Niazi 2022). Essentially, these tasks are
Institute of Soil Mechanics, Foundation Engineering non-trivial and often carried out considering
and Computational Geotechnics, Graz University
empirical interpretation approaches and simplified
of Technology, 8010 Graz, Austria
e-mail: [email protected] assumptions that may have limited applicability in

Vol.: (0123456789)
13
 Geotech Geol Eng

engineering practice. Automating these tasks holds 2 Test Site

promise for enhancing efficiency and transparency.
In this context, this work introduces three dif- The typical soil stratification of valleys in the Alpine
ferent methodologies for the interpretation of CPT region is dominated by postglacial sediments. The
data, with particular focus on the soil stratifica- Salzburg basin was formed during several glacial
tion, fluctuation of CPT readings and soil param- periods. During the post-glacial period, the retreat
eter determination. The applicability of these of glaciers led to the formation of a substantial lake
approaches is demonstrated for a test site located within the basin, mainly filled with fine-grained sedi-
in Salzburg, Austria, detailed in Sect. 2. The core ments. As a result, these young sediments are often
of this paper is divided into three sections, where characterized by a high water table and a normally to
the different CPT data interpretation strategies slightly underconsolidated state (Oberhollenzer et al.
are examined. In Sect. 3, machine learning (ML) 2019, 2021a).
models are used to automatically determine soil The test site is located in Salzburg, Austria. Over
layer boundaries in three-dimensional space. Sec- a rectangular area of approximately 90x50 m, several
tion 4 introduces an intuitive approach to account tests were performed. 6 piezocone penetration tests
for the spatial uncertainty of CPT readings during (CPTu), 4 cone penetration tests (CPT), 4 seismic
the design and update, respectively, of CPT inves- dilatometer tests (SDMT) and 7 core drillings were
tigation layouts. Section 5 illustrates an ongoing executed, as shown in Fig. 1. The location of these
research project focusing on the development of a tests is documented in Fig. 1. The CPTu and CPT
system for automated soil parameter determination. tests were conducted using cones with identical cross-
In this section, this framework is used to compute sectional areas (10 cm2) and vertical measurement
parameters for a fine-grained soil layer identified intervals of 1 cm. The SDMT measurements were
by one of the performed CPTs. Section 6 closes taken every 20 cm, while the shear wave velocity was
with the conclusions of this work, discusses advan- measured every 50 cm.
tages and limitations of the developed CPT data The soil profile can be roughly divided into three
interpretation strategies and points out future lines main layers (Fig. 2). Adopting the specifications doc-
of research. umented in EN ISO 14688-1 (2019), they are clas-
sified as: (i) heterogeneous top layer consisting of

Fig. 1  Test site located in Salzburg, Austria. Note: the position of different tests is measured from point (0,0) located in the lower
left corner of the figure

Vol:. (1234567890)
13
Geotech Geol Eng

Fig. 2  CPTu 5/12

sand-silt mixtures (from 0 to 4 m) and sandy gravel of CPTs with the aim to generate ground models is
(from 4 to 7 m); (ii) sand-silt alterations; (iii) clayey frequently bound to the subjective rationale of the
silt. engineers involved.
In recent years, numerous ML techniques have
been applied to geotechnical engineering problems,
3 Machine Learning Based Strata Identification such as soil liquefaction, the estimation of ultimate
capacity of driven piles and prediction of tunnel
The generation of a site specific ground model that boring machine performance (Kumar et al. 2021;
describes the type and extent of the soil layers as well Samui 2012; Fattahi and Babanouri 2017); addi-
as the subsoil properties is one of the fundamental tional geotechnical application cases of ML models
processes in geotechnical engineering. In traditional are presented in Gomes Correia et al. (2013) and
approaches to this task, the results of available soil Ebid (2021). To reduce subjective influences on the
investigations, such as trenches and core drillings, ground model and ensure a robust and comprehensi-
are usually combined with soundings. Subsequently, ble basis for the planning and execution of geotechni-
individual soil layers and their properties are deter- cal structures, a data-driven strategy allowing for the
mined based on engineering experience. The mini- automatic derivation of ground models (that is, solely
mum number of required core drillings and soundings based on investigation data) is pursued. The under-
is specified in respective standards (e.g., ÖNORM lying database was published by the Institute of Soil
2017). This process requires from the engineers Mechanics, Foundation Engineering and Computa-
involved not only broad expertise in the interpretation tional Geotechnics at Graz University of Technology
of subsurface investigations, but often also specific (TU Graz) in 2021 (Oberhollenzer et al. 2021b).
knowledge of regional characteristics. In many cases,
the latter are only partially substantiated by empirical 3.1 Theoretical Considerations
interpretation schemes, such as Soil Behaviour Type
diagrams (e.g., Robertson 2009, 2010, 2016) or ana- The main objective of the CPT data interpretation
lytical correlations. Consequently, the interpretation strategy presented in this section is to automatically

Vol.: (0123456789)
13
 Geotech Geol Eng

determine soil layer boundaries in three-dimensional 3.2 Data

space. For this purpose, the CPT readings are first
classified in accordance with the European standard The database used in the present study was published
(EN ISO 14688-1 2019) employing a tailor-made by the Institute of Soil Mechanics, Foundation Engi-
ML model that accounts for the dominant grain size neering and Computational Geotechnics (Oberhol-
(gravel, sand, silt, clay, organics). As a key result of lenzer et al. 2021b) and can be downloaded from the
the performed classification, the ground model is sub- institute’s homepage following the subsequent link:
divided into three categories (fine-grained, coarse- https://​ w ww.​ t ugraz.​ a t/​ i nsti​ t ute/​ i bg/​ forsc​ h ung/​
grained, transitional). Empirical thresholds recovered numer​ische-​geote​chnik/​daten​bank/
from multiple tests are used to differentiate between The dataset consists of 1339 CPTs sampled at dif-
the individual classes considering their probability ferent test sites in Austria and Germany. 490 tests
distributions. For example, a data point that is clas- were additionally supplemented with soil class labels
sified as clay with 80% probability is assigned to the according to EN ISO 14688 that were captured using
category “fine-grained”. Ambiguous points for which core drillings. Based on these core drillings, Ober-
no clear distinction can be made (e.g., sand-silt-clay hollenzer et al. (2021b) defined seven classes rang-
mixtures) are classified as “transitional”. It is pointed ing from coarse to fine-grained soils. In this context,
out that the reduction from originally five soil classes first studies focusing on the soil classification with
to three categories smears the CPT data interpreta- ML models showed promising results (Rauter and
tion to a certain degree, that is, similar soil types are Tschuchnigg 2021). In the next subsection, a novel
grouped together leading to a reduction of fluctua- approach is proposed where the soil domain is clas-
tions in the soil layer predictions. Research work con- sified with regard to the EN classes (EN ISO 14688-1
cerning this aspect is part of ongoing research. 2019).
Since the transition between individual soil lay-
ers does not occur at distinct points, but over verti- 3.3 Classification Models
cal intervals, it is necessary to define a process which
automatically determines distinct points represent- The classification scheme employs the “Oberhollen-
ing the soil layer boundaries. For this task, a Kernel zer classes” first presented in Rauter and Tschuchnigg
Density Estimation (KDE) (Pedregosa et al. 2021) (2021). These classes are specified using a random
technique is deployed. Specifically, a KDE graph is forest classifier built on the well-established scikit-
created for each predicted soil class (transitional, learn Python library (Pedregosa et al. 2011). Table 1
coarse-grained and fine-grained). Soil layer bounda- provides selected statistical metrics of the underlying
ries are subsequently approximated at the intersec- database. The database comprises 1,025,284 sam-
tions of the dominant graphs. The advantage of this ples with grain-size-based soil class labels (“Ober-
method is, on the one hand, that soil layer boundaries hollenzer classes”) and 2,514,262 samples classified
can be described by explicit points, on the other hand, by means of soil behaviour types (SBT, SBTn, Mod.
this approach suppresses tedious fluctuations and SBTn); for details, the interested reader may refer
noise inherent to CPT readings. to Rauter and Tschuchnigg (2021). As a key differ-
In combination with the positional data of the CPT ence, modified target classes are considered; instead
test locations, the classification results can be used to of the original seven "Oberhollenzer_classes", the
generate 3D ground models, in which the soil layers present classification scheme is based on the domi-
mark the boundaries of the predicted soil classes that nant soil type (Gravel (Gr), Sand (Sa), Silt (Si), Clay
occupy similar properties. In a subsequent step, the (Cl) and Organic (Or)). The ML model was built
ground model may serve as valuable basis for numer- using 100 estimators with a maximum depth of 18,
ical simulations (e.g., FE Model), including reliable which means that there are 100 decision trees in the
soil constitutive parameters. However, the develop- Random Forest with a maximum of 18 splits. With
ment of a linkage between the derived 3D ground regard to the classification performance, the ML
model and finite element codes is part of ongoing model achieved a value of 0.81 for accuracy, recall,
research, hence beyond the scope of the present study. precision and f1-score, which is slightly higher than

Vol:. (1234567890)
13
Geotech Geol Eng

Table 1  Selected metrics Target Feature Mean Standard deviation Min Max
describing database detailed
in Rauter and Tschuchnigg Oberhollenzer_classes Depth (m) 13.22 10.58 0.01 75.92
(2021)
qc (MPa) 5.34 8.41 −8.61 101.73
fs (MPa) 54.76 70.40 −99.90 1591.40
Rf (%) 2.49 38.16 −100.00 22,000.00
𝜎v (kPa) 251.12 200.94 0.19 1442.48
u0 (kPa) 122.54 103.23 0.00 744.48
𝜎 v ′ (kPa) 128.58 99.69 0.09 697.70
Soil behavior types Depth (m) 12.40 10.43 0.01 103.00
(SBT, SBTn, ModS- qc (MPa) 5.57 8.48 −8.61 122.90
BTn)
fs (MPa) 64.56 254.20 −100.00 47,436.00
Rf (%) 2.70 43.32 −100.00 30,000.00
𝜎v (kPa) 235.68 198.21 0.19 1957.00
u0 (kPa) 114.86 101.48 0.00 1010.43
𝜎 v ′ (kPa) 120.83 98.62 0.09 946.57

the results obtained with the original "Oberhollenzer_ parametric studies with the aim to ensure practica-
classes" (0.75). ble classification results. However, it is explicitly
stated that the optimization of these parameters is
3.4 Machine Learning Methodology part of ongoing research.
The classification scheme described in the above
In analogy to CPT data interpretation strategies paragraph serves as basis for the automatic determi-
based on “Oberhollenzer classes” or Soil Behaviour nation of soil layer boundaries. The density distribu-
Type Charts, noise may have a detrimental effect on tions of the respective categories are calculated and
the ML-based prediction of EN classes. In order to displayed on a chart using the Kernel Density Esti-
overcome this problem in the course of the soil layer mation (KDE) model employed in the scikit-learn
boundary detection a novel classification approach library (Pedregosa et al. 2021). In the present case, a
has been developed. Subsequently, key aspects of the "Gaussian" kernel with a bandwidth of 0.9 was found
underlying workflow are presented. suitable for this task. The dominant distribution in
As a first step, the CPT measurements are clas- each case describes the predominant category; a soil
sified using the ML model. Based on the predicted layer boundary is therefore recovered at the intersec-
class probabilities, the samples are divided into tion point of two dominant distribution curves. This
three categories, namely coarse-grained, fine- approach has the advantage that the noise in the data
grained and transitional. It should be pointed out is considerably suppressed. The disadvantage, how-
that the transformation of EN classes into these cat- ever, is that thin layers may be ignored; moreover, a
egories allows for the direct identification of soil very uneven distribution of categories may lead to
layer boundaries. The class probabilities are used inaccurate positions of the predicted soil layer bound-
to categorize the classified CPT measuring points aries. The methodological steps can be summarized
on the basis of empirical threshold values. In more as follows (Rauter and Tschuchnigg 2022):
detail, a sample is classified as “coarse-grained” if
the proportional sum of gravel and sand exceeds 1. For each EN class, the determinant soil class is
80%; inversely, the measuring points are regarded as evaluated. For example, for siSa (si = silt, Sa =
"fine-grained" if (1) the proportion of clay and silt sand), sand would be identified as dominant soil
exceeds 65% and (2) the proportion of gravel and class as well as the corresponding target class
sand is less than 20%. Points for which no distinc- with respect to the ML model.
tion can be made are categorized as "transitional". 2. A classification model based on a random forest
The threshold values were defined on the basis of algorithm is built and trained using the input fea-

Vol.: (0123456789)
13
 Geotech Geol Eng

tures (qc, Rf , depth) and the new targets (gravel, 5. The soil class density distributions computed
sand, silt, clay, organics). with the KDEs are plotted in one graph, whereas
3. In the scope of the predicted soil class labels, the curves with the highest magnitude are locally
class probability is of particular interest as it is considered as prevailing soil type. Intersections
used to distinguish between the fine-grained, of the distribution curves represent soil layer
coarse-grained or transitional soil labels. For boundaries.
example, if the sum of the probabilities for sand
and gravel is greater than the sum of the probabil- The interested reader may refer to Rauter and
ities for clay and silt (e.g., 80/20 ratio), the sam- Tschuchnigg (2022) for more detailed information
ple would be classified as coarse-grained, oth- about the ML-based methodology, in combination
erwise as fine-grained, and if no clear statement with a simple benchmark test.
can be made, the sample would be classified as Figure 3 visualizes the presented workflow. The
transitional soil. first plot shows the CPT measurement data ( Rf and
4. The identified soil classes (coarse-grained, fine- qc). The second plot indicates the classification
grained and transitional) are displayed by means according to Robertson’s normalized Soil Behaviour
of a binary plot, where each CPT recording Type chart (Robertson 2009), which serves as refer-
is associated with a unique soil class (coarse- ence basis for comparison with the EN-based classi-
grained, fine-grained and transitional). Based fication. The third plot yields the probability distri-
on this distribution, a kernel density estimation bution of the EN classes. The fourth plot shows the
is performed for each class. It is found that a binary plot with distinct soil class labels in vertical
Gaussian kernel with a bandwidth = 0.9 (that is, direction (coarse-grained, fine-grained and transi-
a hyperparameter that controls the smoothness of tional). The fifth plot depicts the probability density
the class probability curves) is reasonable to pre- distribution of the three soil classes. In the present
dict realistic soil layer boundaries. case, two intersections corresponding to two soil

Fig. 3  Example for the presented workflow. Plot of the CPT data, SBT classification, class probabilities of EN-classification, clas-
sification in the three newly introduced categories and density distribution (f.l.t.r.)

Vol:. (1234567890)
13
Geotech Geol Eng

Fig. 4  Cross-section A-A. Comparison between predicted soil layer boundaries and reference solution (dashed line)

Fig. 5  Cross-section B-B. Comparison between predicted soil layer boundaries and reference solution (dashed line)

layer boundaries can be identified (black dash-dotted defined by engineers in the course of a construction
lines). project. In addition to the CPTs, core drillings and
In the next subsection, the presented workflow is seismic dilatometer tests (SDMT) served as a basis
applied to seven different CPTs to identify the soil for soil layer determination in the reference model,
layer boundaries of the test site described in Sect. 2. which will be referred to as reference solution or ref-
For validation purposes, a comparison with the engi- erence boundaries.
neering ground model generated in the course of the Figures 4b and 5b show the spatial distribution of
construction project is provided as well. the predicted soil layers from two different perspec-
tives, in which the black vertical lines show the posi-
3.5 Results tion of the individual CPTs, the red surfaces mark the
top of the ground model, the green surfaces indicate
The ground model under study follows the dimen- the layer boundaries between coarse-grained and tran-
sions of the rectangular test site (approximately sitional soil, and the blue surfaces illustrate the tran-
90x50 m and a depth of 40 m). The arrangement sition from transitional to fine-grained soil. The sur-
of the CPTs in the ground model corresponds to faces were recovered employing the widely-applied
the actual arrangement of the soundings in the field surface triangulation interpolation technique; com-
(see Fig. 1). The sections used for comparison were pare Felic (2023).

Vol.: (0123456789)
13
 Geotech Geol Eng

In the next step, the predicted soil layers are com- utilize an appropriate database that shall be continu-
pared to the reference solution along two selected ver- ously maintained to ensure a satisfactory performance
tical cross-sections. For the purpose of comparison, of the presented strategy in future applications. Fur-
the locations of these cross-sections are based on the thermore, the hyperparameter set employed in the ML
sectioning originally established during the design of and KDE models have to be evaluated and updated as
the original construction project. The first cross-sec- soon as the database is extended with new samples.
tion extends along CPTu 1/12, 2/12 and 3/12, and is This aspect is part of ongoing research at the Insti-
illustrated in Fig. 4a. The coloured areas represent the tute of Soil Mechanics, Foundation Engineering and
individual layers according to the three categories, Computational Geotechnics.
i.e., coarse-grained, fine-grained and transitional. The
dashed lines mark the soil layer boundaries of the
reference solution (as used for the design). As can be 4 Dynamic Updating Strategy for CPT Layouts
observed from the results, a very good agreement of
the predicted soil layer boundaries and the reference The spatial interpretation of CPT measurements, for
solution is achieved. The maximum deviation from example, on the basis of ML techniques as discussed
the reference solution is around 1 m. in Sect. 3, is sensitive to the CPT investigation layout
The second cross-section used for comparison (Jaksa et al. 2005b). Since the spatial interpretation is
extends along CPTu 2/12, 5/12 and CPT 1/15. (Note: deterministic at measured locations, it is unknown at
CPT 1/15 and CPTu x/12 were executed at different unmeasured locations and needs to be assessed based
times and from non-identical levels). Results con- on predictions from the available limited number of
cerning cross-section B-B are shown in Fig. 5a. Simi- measurements (Kulatilake and Um 2003). In many
lar to cross-section A-A, the soil layer boundary at the instances, the accuracy of predictions increases as the
interface between “transitional” and “fine-grained” scope of investigations and available data increases.
agrees reasonably well with the reference solution. However, narrow-spaced CPT layouts parallel both
However, the soil layer boundary at the interface high costs spent in carrying out the investigations and
between “coarse-grained” and “transitional” shows redundant information. Moreover, since every CPT
considerable deviations from the reference solution, leaves a zone of permanently disturbed soil where
especially close to CPT 1/15. soil properties could be considerably different from
the original in-situ state, there is need to account for
3.6 Applicability in Engineering Practice a minimum spacing between adjacent CPT locations
(Al-Sammarraie et al. 2022). Consequently, the selec-
The presented methodology for data-driven and auto- tion of the optimal location and number of CPT tests
matic identification of soil layer boundaries based on represents a multivariate optimization problem that
CPT measurement data represents an attractive alter- may be regarded as trade-off between stratigraphic
native to classical filter algorithms that are currently uncertainty and site investigation costs; cf. Kahlström
used for the determination of soil layers boundaries. et al. (2021).
Comparative analyses presented in this section indi- In common practice, the scope of the geotechni-
cate a good agreement of the predicted soil layer cal site investigations rarely addresses the antici-
boundaries with the reference solutions. However, pated variability of the ground, but it is rather chosen
several open research aspects deserve attention. First, to reduce initial costs. It is widely accepted that this
the performance of ML models used for geotechni- cost-driven approach is likely to result in cost-over-
cal problems is significantly controlled by the qual- runs, for example, due to construction delays invoked
ity of training samples. In addition, it relies on the by unforeseen conditions or uneconomic designs
amount of data and does not extrapolate well beyond resulting from inadequate stratigraphic models (Maz-
the scope of provided training patterns. There is also zoccola et al. 1997; Goldsworthy et al. 2004; Jaksa
a need to develop tailored machine learning methods et al. 2005b). The above discussion inevitably raises
that can account for uncertainty in the global data, the following question (Xie et al. 2022): How to use
variability between sites, and realistic sample sizes a limited budget to estimate the ground variability
(Phoon and Zhang 2023). It is therefore crucial to with the lowermost uncertainty? This section aims to

Vol:. (1234567890)
13
Geotech Geol Eng

contribute towards answering this question by means As indicated by Fig. 6a, 𝜙(x)-values close to 1
of an intuitive strategy that allows for an increase of signify a high degree of spatial variability in related
the ratio between gained understanding of the spatial soil profiles (for example, due to the nugget effect
subsoil variability and costs spent for carrying out (Jaksa et al. 2005a)), whereas 𝜙(x)-values slightly
CPTs. The applicability of this method is demon- above 0 indicate regions of nominally homogeneous
strated based on both synthetic and real test cases. soil properties. In simple terms, 𝜙(x) allows to iden-
tify spatial regions of low correlation between neigh-
4.1 Methodology bouring CPT datasets, thereby taking into account
the proximity of unsampled points to available CPT
Unlike conventional methods that traditionally con- locations by means of a deterministic spatial weight-
cern the ground variability solely based on the dis- ing approach. To prevent the execution of CPTs
tance between CPTs, such as kriging-based (Krige that result in redundant information being obtained,
1951) or distance-from-nearest-information (Weil the CPT layout should be sequentially densified in
2020) sources, the presented framework combines regions with high 𝜙(x)-values. With reference to 6a,
geotechnical and spatial sampling information to carrying out a new CPT at x1 would yield valuable
reduce the stratigraphic uncertainty in the field where insight to the soil stratification, whereas additional
CPT measurements are conducted. In more detail, the information gained from placing an additional CPT at
proposed strategy characterizes the 3D stratigraphic x2 is expected to be relatively limited. From a practi-
model based on the dimensionless scalar-valued field cal perspective, this enables to gain maximum insight
quantity termed level of fluctuation 𝜙(x) ∈ [0, 1] (not to the 3D stratigraphic model at minimum costs.
be confused with the scale of fluctuation occupying Nevertheless, it must be noted that 𝜙(x) introduces a
the unit m (Vanmarcke 1977)). In this respect, 𝜙(x) relative measure that helps to detect local regions of
embodies an unbiased estimator capable of character- high spatial variability on a site-specific level only; as
izing the spatial variability of the CPT measurements. a consequence, quantifying generally applicable 𝜙(x)

Fig. 6  (a) Schematic description of scalar-valued field quantity termed level of fluctuation 𝜙(x) and (b) framework of proposed
model

Vol.: (0123456789)
13
 Geotech Geol Eng

-limits that can readily be adopted to any site investi- number of CPTs considered, the vertical coordinates
gation lies beyond its current capabilities. z, zmin and zmax are described in Fig. 6a. Equation 1
Figure 6b illustrates the process of generating the is numerically integrated using the trapezoidal New-
level of fluctuation which has been implemented in ton-Cotes rule with a step size equal to the CPT pen-
an object-oriented framework using the programming etration interval, traditionally h = 0.01 m. According
language Python. It is worth mentioning that the pre- to this nomenclature, L2-norm matrix entries A(i,j)
sented framework is theoretically applicable to any describe the deviation of qc-profiles obtained from the
continuously sampled 1D quantity, such as the sleeve ith and jth CPT recording, respectively. As diagonal
friction fs, friction ratio Rf or cone tip resistance qc. terms (i = j ) concern the inverse correlation between
For the sake of clarity, subsequent considerations are identical qc-profiles, the corresponding diagonal fill-
restricted to qc-samplings since the latter is frequently ins reduce to zero (i.e., they are perfectly correlated;
used as critical parameter in the CPT-based pile design see Fig. 7a).
(Bustamante and Gianesselli 1982; Hamza and Bellis Step 3: Partition of stratigraphic model
2016; Abu-Farsakh and Shoaib 2024). The details of As indicated in Fig. 7b, a regular tessellation pat-
this procedure can be summarized as follows: tern is used to generate a quadratic grid. The level
Step 1: Loading and pre-processing of CPT data of fluctuation 𝜙 is evaluated at cell-centered discrete
The one-dimensional qc-profiles obtained from points xp ∈ ℝ2. As suggested by previous researchers
CPT field tests are simultaneously loaded and cleaned. in view of similar problems (Jaksa et al. 2005b; Xie
The latter is realized by means of the 1D Gaussian fil- et al. 2022), it is recommended to select a minimum
ter technique, similar to Liu et al. (2021). This allows grid size of 0.5 x 0.5 m. This ensures a convenient
to minimize spurious contributions to the calculation resolution of the continuous field quantity by means
of 𝜙(x) originating from noise and outliers. of discrete 𝜙(xp )-values, as well as a high level of
Step 2: L2-norm matrix computational efficiency considering the large scope
The inverse degrees of correlation between indi- of many geotechnical fields.
vidual pairs of qc-profiles, q(CPTi)
c (z) and qc
(CPTj)
(z), are Step 4: Level of fluctuation
quantified through the scalar-valued entries A(i,j) of The discrete evaluation of 𝜙 at all cell-centered
the symmetric L2-norm matrix A ∈ ℝn×n: grid points xp forms the core of the multi-step pro-
(CPTj )
cedure and comprises several sub-steps; see Fig. 8a.
A(i,j) =‖q(CPT
c
i ) (z) − q
c (z)‖L2 (zmin ,zmax ) The first sub-step concerns the evaluation of the dis-
� tance vector d ∈ ℝn which contains the horizontal
zmax (1)
∫zmin
length between xp and all CPT positions; see Fig. 8b.
(CPTi ) (CPTj )
= [qc (z) − qc (z)]2 dz
The calculated values are stored in the form of an
attribute table at all xp, similar to Fig. 7a. In the next
In Eq. 1, i, j ∈ [1, n] denote CPT index numbers,
such as considered in Fig. 7b, where n is the total

Fig. 7  (a) Numbering system of L2-norm matrix A and (b) corresponding CPT layout

Vol:. (1234567890)
13
Geotech Geol Eng

Fig. 8  (a) Sub-steps required to evaluate 𝜙(xi ) and (b) point-wise evaluation of distance vector d

sub-step, d is used for a point-wise evaluation of the distributed soil properties where further investiga-
weight matrix W(xp ) ∈ ℝn×n, formally written as: tion is supposed to yield redundant information; cf.
Fig. 6.
1∕[di (xp ) ⋅ dj (xp )]2 It should be pointed out that the CPT layout opti-
W(xp )(i,j) = ∑n ∑n (2)
i=1 j=1 1∕[di (xp ) ⋅ dj (xp )]
2 mization procedure presented in this subsection
can readily be extended to analyse the ground vari-
In Eq. 2, the weight matrix entries W(xp )(i,j) account ability in the third dimension (vertical z-direction)
for the proximity of xp to the locations of considered as well; for this purpose, the level of fluctuation is
CPT pairs, and satisfy the condition of an unbiased alternatively computed at voxel-centered discrete
∑n ∑n
estimator i=1 j=1
W (i,j) = 1. As formulated in points xp ∈ ℝ3. This allows, for example, to antici-
Eq. 3, the final sub-step links contributions originat- pate potential consequences of a CPT investigation
ing from geotechnical data A(i,j) with spatial sample depth reduction, or to cross-check whether FE sub-
information Wp (i,j) in the form of the level of fluctua- domains experiencing high pile load concentrations
tion 𝜙(xp ): are circumscribed by reliable site investigation data.
∑n ∑n (i,j)
From a computational point of view, this requires
i=1 j=1 A ⋅ W (i,j) the consecutive application of the model framework
𝜙(xp ) = �∑ ∑
n n (i,j) (i,j)
� (3) described in Fig. 6b on multiple equally spaced lay-
max A ⋅ W
i=1 j=1
ers with constant depth interval, instead of a single
formation spanning the entire CPT investigation
The normalizing denominator component depth, see Fig. 6a. As a consequence, the integra-
involves a loop over all discrete points and facili- tion limits ( zmin , zmax ) in Eq. 1 have to be replaced
tates a systematic quantification of the sought field by corresponding pairs of interval boundaries,
quantity in a range between 0 and 1. whereas the mean value representing the z-coordi-
Step 5: Visualization and interpretation nate is additionally stored at all xp . Moreover, since
The level of flucation at discrete data points the normalization is now enforced on xp occupied
𝜙(xp ) is effectively visualized using standard trian- by different layers, the normalizing denomina-
gular irregular networks in combination with 2D tor component in Eq. 3 has to be equipped with an
filled contours. On a site-specific scale, this allows additional loop over all layers, each of which incor-
to anticipate regions with high spatial variability porates a set of xp where the three-dimensional level
inside the 3D stratigraphic model, which can be of fluctuation 𝜙3D (xp ) ∈ [0, 1] is explicitly com-
mathematically described by 𝜙(xp ) → 1−. Inversely, puted. For the sake of clarity, corresponding figures
𝜙(xp ) → 0+ characterize areas of homogeneously explaining both strategies (𝜙 , 𝜙3D ) are reported in
the subsequent Sects. 4.2 and 4.3.

Vol.: (0123456789)
13
 Geotech Geol Eng

4.2 Synthetic Case Application result in redundant information being obtained. In

this context, the zero-valued L2-norm matrix listed in
To demonstrate the capabilities of the procedure Fig. 10a confirms the implementational fidelity of the
described in Sect. 4.1, a total of three synthetic CPT code.
layouts have been examined; see Fig. 9a−c. The latter Unlike the homogeneous case, the cross-symmet-
incorporate the measurements of four CPTs (Fig. 9d) ric as well as the heterogeneous CPT layout yield
which are located at the corners of a squared 100 m x non-uniform estimates of the qc-variability across
100 m site, and adopted from the test site described the synthetic test sites. In both cases, the CPT corner
in Sect. 2. The CPT configurations comply in terms points are circumscribed by 𝜙(xp )-regions close to 0,
of spatial arrangements, but differ with regard to the which could be expected due to the close proximity
distribution of qc-profiles. It is evident from Fig. 9a to at least one CPT position. However, considerable
that the homogeneous case (i.e., identical qc-profiles deviations between both synthetic cases are observed
at all four corner points) yields 𝜙 = 0 ∀ xp, indicat- at greater distances to the corners: While the cross-
ing a homogeneous distribution of qc across the test symmetric layout results in a radial-symmetric pat-
site; hence, carrying out additional CPTs is likely to tern, where the peak values are aligned with the

Fig. 9  Site investigation schemes and corresponding 𝜙(xp )-distributions obtained with (a) homogeneous, (b) cross-symmetric and
(c) heterogeneous case; (d) CPT data considered

Fig. 10  L2-norm matrices obtained with (a) homogeneous and (b) heterogeneous CPT layout

Vol:. (1234567890)
13
Geotech Geol Eng

vertical and horizontal site axis, peak-valued 𝜙(xp )- Unlike the synthetic test cases investigated in
areas evaluated for the heterogeneous test case are (as Sect. 4.2, 𝜙-values are also calculated at spatial
expected) non-symmetric. Obviously, they are shifted points xp that are located outside the convex hull-
towards the site edge connecting CPT 4 and CPT 1, polygon; i.e., the connection line of outermost CPT
or more specifically, the site edge between the CPT samplings (Fig. 11a). At a first glance, the extrapo-
pair occupying the maximum L2-norm matrix entry, lated 𝜙-regions appear consistent with their interpo-
see Fig. 10b. Again, this observation replicates the lated counterparts in close proximity to the convex
anticipated pattern, and confirms the reliability of the hull-polygon. However, due to the lack of physical
implemented workflow. evidence in extrapolated regions the use of related
results inherently implies a certain degree of ran-
4.3 Real Case Application domness (Henke et al. 2020); hence, care should be
taken when using extrapolated 𝜙-values for CPT lay-
In this subsection, the proposed strategy is employed out optimization purposes. As expected, minimum 𝜙
to identify both the optimal CPT location and a rea- -values are obtained in close proximity to the position
sonable investigation depth to supplement CPT sam- of recorded CPT samplings. This observation fol-
pling data described in Sect. 2. The corresponding lows the same pattern as discussed in Sect. 4.2 and
qc-recordings obtained from 10 CPTs are plotted in showcases the influence of the weight matrix W ; see
Fig. 11b. In accordance with Sect. 3.5, the adopted Sect. 4.1. On the contrary, Fig. 11a predicts the maxi-
local coordinates comprise a vertical z-axis starting mum level of fluctuation at coordinates (x = 20 m,
at the basement level. In this respect, the investiga- y = 36 m); in turn, the latter represents the optimal
tion area includes the subsoil formation extending position for the execution of subsequent CPT testing
between 0 m and -30 m of the test site, with at least in order to acquire maximum insight to the natural
two relevant soil layer boundaries; cf. Sect. 3. variability of qc, respectively, at minimum costs.

Fig. 11  (a) Site investigation layout and corresponding 𝜙(xp )-contour plot, as well as (b) 𝜙3D (xp )-contour plots presented in y–z and
x–y plane, respectively; (c) CPT data considered

Vol.: (0123456789)
13
 Geotech Geol Eng

The above observations incorporate information early design phase, the ability to effectively com-
on the horizontal variability of qc, but only limited municate geotechnical information and risk associ-
information on its vertical distribution. Nevertheless, ated with the 3D stratigraphic model to all project
soil properties are typically distributed with greater stakeholders including authorities, architects, cli-
variation in the vertical than in the horizontal direc- ents, contractors and engineers is a central challenge
tion, primarily due to geological processes that form for geotechnical engineers (Kahlström et al. 2021).
soils (Jaksa et al. 2005b); for example, this tendency In this context, the proposed strategy may help geo-
can be triggered by changes in the porewater chemis- technical engineers to meet this demand. In principal,
try and groundwater levels, the influence of climatic the authors have identified three use cases where the
effects or the diagenesis; cf. Graham and Shields presented strategy may extend the toolbox of geotech-
(1985). Consequently, it may be important to account nical engineers: (1) objectification and prioritisation
for the vertical ground variability when updating the of CPT layout design, (2) effective communication
CPT layout. The incorporation of this aspect also of geo-spatial uncertainty to project stakeholders
facilitates a data-driven determination of the CPT through the power of visualization and (3) quality
investigation depth on the basic premise of uncer- assurance of CPT measurements (note that peak 𝜙
tainty reduction, rather than randomness. /𝜙3D-values may also be triggered by measurement
As theoretically explained at Sect. 4.1, the method- errors).
ical framework can readily be extended to allow for a In view of the current trend to integrate 3D strati-
three-dimensional analysis of the test site soil domain graphic models to the Building Information Modeling
on the basis of 𝜙3D (xp ). From a geometrical point of (BIM) process (Henke and Lerch 2020; Giangiulio
view, the only modification in the presented workflow et al. 2022), 𝜙/𝜙3D-values may be stored as soil vol-
for calculating the level of fluctuation concerns the ume attribute. In this way, they can be readily used
partition of the stratigraphic model into cubic voxels, as relative measure that link the level of uncertainty
instead of two-dimensional quadratic grid elements; with geotechnical parameters listed in respective soil
see step 3 in Sect. 4.1. Consequently, 𝜙3D-values are volume attribute tables. Moreover, 𝜙/𝜙3D-attributes
calculated at voxel-centered discrete points xp ∈ ℝ3 may be dynamically updated as new CPT samplings
which are defined in all three spatial dimensions. As are performed using the framework documented in
depicted in Fig. 11a, the abundant data is effectively Fig. 6b.
visualized by means of perpendicular cross-sections For the sake of clarity, the demonstration cases
presented in both the x–z and y–z-plane, respectively. presented above assess the geotechnical contribution
Obviously, peak 𝜙3D-values indicating a high verti- to 𝜙/𝜙3D-quantities solely based on qc-recordings (sin-
cal variability occur in the near-surface whereas soil gle-factor approach). Depending on the site investiga-
regions at vertical distances greater than 6 m below tion objectives and the soil properties, it may be ben-
the basement are dominated by fairly homogene- eficial to replace Eq. 1 by a weighted-averaging
ous geotechnical conditions. From a practical point scheme, in which the L2 −norm matrix entries A(i,j)
of view, subsequent site investigation effort should subsume contributions stemming from the cone tip
therefore focus on the near-surface subsoil instead resistance A(i,j)
q
, the sleeve friction Af and the fric-
(i,j)

of deep soil locations. In some instances, this could

c s

tion ratio AR . A validation of the multi-factor

(i,j)
encourage the application of low-cost CPT-alterna- f

tives, which could be a conceivable alternative for approach, however, is beyond the scope of this paper,
site investigation in near-surface regions, cf. Palla and subject of ongoing research.
et al. (2008).
5 An Automated System for Determining
4.4 Generalisation and Employment in Practice
Constitutive Model Parameters
It is widely acknowledged that, in building projects
Numerical analyses have several advantages com-
and civil engineering, the main source of financial
pared to traditional methods. One of the main advan-
and technical risk lies in the ground (Jaksa et al.
tages is the level of detail that could be obtained in
2005b; Rana et al. 2023). In particular during the

Vol:. (1234567890)
13
Geotech Geol Eng

several geotechnical engineering problems, such as an adaptable parameter determination framework.

soil-structure interaction (Brinkgreve 2019). The suc- Transparency is ensured by describing how available
cess of numerical analysis is influenced by many fac- information is used to reach a certain solution, while
tors. One of the key factors is determining the con- adaptability is achieved by allowing the users of the
stitutive model parameters properly. It is usually the system to incorporate their knowledge and experience
case that these parameters are determined based on into the system. Van Berkom et al. (2022) illustrated
limited soil data. Most often, these parameters need the graph-based approach and presented an example
to be assessed based on experimental tests (e.g., triax- of the determination of parameters for coarse-grained
ial and oedometer tests) that are not always available soils based on CPT data. Marzouk et al. (2022)
in projects (especially not in an early design phase). extended the framework by including parameters and
In-situ tests offer an alternative method to deter- empirical relationships for fine-grained soils. The
mine soil parameters. The cone penetration test extension of the system to include other in-situ tests
(CPT) is one of the most popular in-situ tests as it is is in progress, where the dilatometer (DMT) has been
quick and can be used in soil profiling and determin- added to the framework.
ing soil parameters (Lunne et al. 1997; Ricceri et al. This section presents a case study where the APD
2002; Anagnostopoulos et al. 2003; Schnaid 2009). system is used to determine soil parameters based
Furthermore, compared to laboratory tests, CPT has on CPT data that were obtained from the test site
other advantages, including lower cost and minimal described in Sect. 2. Section 5.1 briefly describes the
disturbance of the soil. Contrary to laboratory tests, APD framework, while Sect. 5.2 presents selected
it is not possible to determine soil parameters directly empirical relationships used to determine soil param-
from the results of in-situ tests. However, soil param- eters. In Sect. 5.3, the output of the APD for a CPT is
eters can be obtained from in-situ test results based on illustrated. The conclusions of the study are summa-
empirical relationships. It is often the case that sev- rized in Sect. 5.4.
eral relationships exist to determine the same parame-
ter, presumably leading to a (wide) range of values for 5.1 Automated Parameter Determination (APD)
the parameters of interest. This variation is attributed Framework
to the applicability of these relationships, as they are
not suitable for all situations (e.g., specific soil types). 5.1.1 Framework
In literature, there are some guidelines focusing on
the interpretation of soil parameters from CPTs (e.g., Five connected modules form the framework. A
Kulhawy and Mayne 1990; Lunne et al. 1997; Mayne schematic representation of the modules is shown in
2014; Robertson 2015). Fig. 12. The first module (GEF reader) imports CPT
An ongoing research project aims to create an raw data in Geotechnical Exchange Format (GEF). In
automated parameter determination (APD) frame- the next step, the CPT measurements (cone resistance
work to determine constitutive model parameters qc, sleeve friction fs and porewater pressure readings
from in-situ tests. This system is based on a graph- u2) are transferred to the second module (CPT layer
based approach that relies on graph theory princi- interpretation). In this module, the soil behaviour
pals, as will be detailed in Sect. 5.1. The main moti- type (SBT) is determined based on Robertson’s modi-
vation of this project is to create a transparent and fied non-normalized SBT chart (Robertson 2010).

Fig. 12  Schematic repre-

sentation of the parameter
determination modules

Vol.: (0123456789)
13
 Geotech Geol Eng

Nevertheless, other SBT charts (such as Robertson’s Table 2  SBT zones Robertson (2010)
normalized chart (Robertson 2009) and Robertson’s Zone Soil behaviour type (SBT)
updated normalized chart (Robertson 2016)) could
be used as well. Afterwards, the CPT profile is strati- 1 Sensitive fine-grained
fied into several layers sharing the same SBT. At the 2 Clays—organic soil
moment APD does not use any ML based strata iden- 3 Clays: clay to silty clay
tification (as discussed in Sect. 3). However, the pos- 4 Silt mixtures: clayey silt & Silty clay
sibility of using ML for the soil layer detection (with 5 Sand mixtures: silty sand to sandy silt
SBT charts) will be added to the framework. For each 6 Sands: clean sands to silty sands
layer, the CPT measurements (qc, fs and u2) are aver- 7 Dense sand to gravelly sand
aged in a layer-wise manner. The third module (Layer 8 Stiff sand to clayey sand (overconsolidated)
state) uses the averaged CPT measurements to deter- 9 Stiff fine-grained (overconsolidated)
mine the state of the layers (overconsolidation ratio
OCR and coefficient of earth pressure K0). The fourth
module (Graph-based approach) uses the output of 5.1.3 Graph‑Based Approach
modules 2 and 3 to connect parameters to equations
(correlations) and calculates the parameters. The Graph theory is a branch of discrete mathematics
final module (Constitutive model parameters) con- where relationships between different objects within
verts parameters calculated in module 4 to constitu- a network are studied. The network is described by
tive model parameters. The system is built in the pro- two different objects, namely nodes and edges. Nodes
gramming language Python. describe the entities of the graph, while edges illus-
The layering process and determination of the trate the relationship between two nodes. One of the
"Layer state" are not considered (modules 2 and 3) main advantages of graphs is their ability to repre-
in this contribution. It has to be pointed out that the sent complex systems (e.g., transportation network
layering used in this case study is based on the strati- (Likaj et al. 2013) and social media (Chakraborty
fication obtained from the ML models presented in et al. 2018)). The APD framework uses a weighted
Sect. 3, thus, the implemented layering algorithms directed graph, where the direction between two
have not been used. Furthermore, this section only nodes (sharing a relationship) in the graph is defined
illustrates the determination of soil parameters (out- (Van Berkom et al. 2022). In addition, weights could
put of module 4) without the transition to constitutive be assigned to the edges between the nodes. One-way
model parameters (module 5). roads are an example of weighted directed graphs,
where roads could have an assigned weight represent-
5.1.2 SBT Interpretation ing distance or travel time.
Van Berkom et al. (2022) illustrated the con-
In the present study, the CPT profile is classified cept of the graph-based approach employed in APD
according to Robertson’s modified non-normalized in detail. Figure 13 summarizes the graph-based
SBT chart (Robertson 2010). The chart is based on approach, where source parameters (CPT raw data)
the dimensionless cone resistance (qc ∕pa, where pa are linked via intermediate parameters to destination
is the atmospheric pressure) and friction ratio ( Rf in parameters (soil or constitutive model parameters).
percent, Rf = fs ∕qc 100%). The chart is divided into By using a given set of correlations, the system
nine different zones corresponding to different soil creates all paths (chains of correlations) that link
behaviour types (Table 2). Therefore, fine and coarse- source parameters to the destinations parameters,
grained soils are distinguished by this module. Never- and calculates the values of the destination param-
theless, the classification is not limited to Robertson’s eters from the input values of the source parameters
modified non-normalized SBT chart, as other charts (CPT raw data). Within the framework of APD, the
are already implemented in the system (e.g., Robert- terms ‘correlation’, ‘formula’, ‘equation’, ‘rule of
son’s normalized SBT chart (Robertson 2009) and thumb’ are replaced by the term ‘method’. Param-
Robertson’s updated normalized SBT chart (Robert- eters could be determined based on several ways
son 2016)).

Vol:. (1234567890)
13
Geotech Geol Eng

Fig. 13  Graph-based
approach implemented in
APD

(e.g., tables and charts), therefore this general term a result, method nodes have several incoming edges
is used (Van Berkom et al. 2022). which indicate all the required parameters for this
Searching for a path in a network from one node method (Van Berkom et al. 2022). Methods and
to another is a well-known problem, which is widely parameters that share a relationship must be linked.
used in several applications (Shu-Xi 2012). Sev- For example, a method to compute the coefficient of
eral graph algorithms exist which allow for solving earth pressure according to Jaky (1944) considers
the shortest path problem (finding the shortest path the well-known relationship K0 = 1 − sin(𝜑� ), where
between two nodes in a graph). Nevertheless, graph K0 is the coefficient of earth pressure at rest and
algorithms do not apply to branching paths. Within 𝜑′ denotes the effective internal friction angle of
the parameter determination framework, a path to the the soil. The system must determine the input and
destination node can have more than one source node output for this method (𝜑′ and K0 respectively) and
as the parameters in the path can be obtained from generate links connecting these parameters.
multivariable formulas, which depend on multiple
input parameters (branching paths occur in the frame- 5.1.4 Generating the Graph
work). Consequently, the existing graph algorithms
cannot be applied to the parameter determination As discussed in Sect. 5.1.3, the relationships between
framework (Van Berkom et al. 2022). parameters and methods are defined based on the
To deal with branching paths, two types of nodes output and input(s) of different methods. Parameters
are implemented, namely parameters (green nodes and methods are considered as external inputs to the
in Fig. 14) and methods (blue nodes in Fig. 14). system. Two input files are required, namely, "meth-
Most often method nodes are empirical correla- ods" and "parameters". Due to the adaptability of the
tions that depend on more than one parameter. As framework, users may extend the standard database
of methods and parameters provided by the current

Vol.: (0123456789)
13
 Geotech Geol Eng

version of the system. The system generates links recommended to define the units of all parameters.
between the methods and parameters and calculates The Constraints field is used to apply lower and upper
the intermediate and destination parameters. The two bounds to the parameters. If the calculated value is
different input files used to generate the graph are lower than the lower bound or higher than the upper
provided in comma-separate values (CSV) format bound, it would be discarded. This allows to neglect
(corresponding to parameters and methods). unrealistically high or low values for parameters.
Each CSV file is characterised by special prop- Similar to the reference field in the methods CSV file,
erties. With regard to the methods CSV file, it is the description field is an optional argument, where
required to define the several properties, namely the user could describe the parameter (e.g., OCR is
method_to, formula, parameters_in, parameters_out, the overconsolidation ratio).
validity and reference. The latter should be provided By defining these two CSV files (methods &
by the user in the methods CSV file. Following the parameters), the system imports the two files and cre-
above example concerning the earth pressure at ates links between related methods and parameters
rest coefficient, method_to presents the name of the (parameters_in and parameters_out). The output of
method; in this case it could be method_to_K0. In the module 4 (Graph-based approach) is a graph present-
formula’s field, the corresponding equation should be ing all the links between all the defined parameters
provided, i.e., 1 − sin(𝜙� ). Parameters_in implicitly and methods as well as the calculated values for dif-
defines the input(s) for this method, i.e., 𝜙′. Follow- ferent parameters (e.g., see Fig. 14).
ing the same definition, parameters_out states the
output of the method, i.e., K0. The applicability of
different methods is defined in the validity field. As 5.2 CPT Interpretation
some methods are suitable for all types of soils, other
methods are only valid for coarse-grained soils, while A standard validated database for methods and
others are only applicable for fine-grained soils. parameters has been compiled and is continuously
As presented in Table 2, the SBT is based on Rob- updated and improved. The current version of the
ertson’s modified non normalized SBT chart (Rob- database consists of more than 100 methods and
ertson 2010). Consequently, the validity is (mainly) parameters. Nevertheless, users are responsible for
defined in terms of SBT. If the method is only appli- validating the outcome of the system, even if they
cable to silt, the validity would be SBT(4). Concern- used the provided standard database. They still need
ing the method of coefficient of earth pressure at rest, to apply their experience and knowledge to the out-
the validity would be SBT(1234567). Reference is put of the system. However, with limited geotechnical
an optional field, where the user could mention the knowledge, the system should provide reasonable val-
author of the method (e.g., Jaky_1944). ues for different parameters. However, using all of the
Moving to the parameters CSV file, the following methods in the database could lead to a scatter in the
properties need to be defined: symbol, value, unit, obtained values, which will make the representation
constraints and description. Any parameter that has of the results challenging. To circumvent this problem
been defined in the methods CSV file (in the fields in the present study, graphs are only created based on
of formula, parameters_in and parameters_out) must a selected number of methods and parameters that are
be defined in the parameters CSV file. In the symbol presented as follows:
field, the notation of the parameter (that has been
used in the methods CSV file) is stated (e.g., u for 5.2.1 Initial Parameters
porewater pressure). In the value field, the user could
specify a value for a parameter (e.g., unit weight of An initial estimate of the total (𝜎v) and effective
water). It could also be used by the user to provide (𝜎 v ′) stress is required to compute some of the CPT
"known" values as direct "input" for some param- parameters (e.g., normalized cone resistance). Con-
eters. Unit is an optional field, where the user could sequently, the unit weight needs to be assessed at
provide the unit for all parameters. In order to avoid the beginning of the analysis. In the considered case
problems originating from unit conversion (e.g., using study the initial unit weight is computed according to
qc in MPa in a method that requires qc in kPa), it is Robertson and Cabal (2010):

Vol:. (1234567890)
13
Geotech Geol Eng

𝛾t = 𝛾w [0.27[log Rf ] + 0.36[log (qt ∕pa )] + 1.236] (4) In the following Sects. 5.2.1, 5.2.2, 5.2.3 and 5.2.4,
the selected methods and parameters are presented.
where 𝛾w is the unit weight of water and qt is These methods and parameters serve as basis for gen-
the corrected cone resistance, i.e., defined as erating the graphs in module 4, as will be shown in
qt = qc + (1 − a) × u2, where a is the cone tip net area Sect. 5.3. As mentioned above, in total more than 100
ratio. methods are implemented in the APD framework.
The porewater pressure (u0) is calculated based on
the ground water level (GWL). The GWL is provided
5.2.2 Stress History
in the CPT GEF file, otherwise it could be speci-
fied by the user. Based on the previous information,
Stress history is often represented by the overconsoli-
the total and effective stress are computed. As a key 𝜎p�
result, the following “CPT parameters” are computed: dation ratio defined as OCR = 𝜎v�
, where 𝜎p′ is the ver-
tical preconsolidation stress. Numerous methods are
• Normalized cone resistance available in the database to compute OCR. In this
q t − 𝜎v study, two methods were selected to determine OCR
Qt = (5) as follows:
𝜎v�
𝜎p′ 0.33(qt − 𝜎v )m
′

• Normalized porewater pressure OCR = = (10)

𝜎v′ 𝜎v′
(u2 − u0 ) 1.
Bq = (6) by Mayne et al. (2009), where m′ is the yield
(qt − 𝜎v )
stress exponent that increases with fines content
• Normalized cone parameter with variable stress and decreases with mean grain size. This corre-
exponent n that varies with soil behaviour type lation is valid for different types of soils. Mayne
index ( Icn) and is calculated in an iterative man- (2017) proposed determining m′ from CPT mate-
ner: rial index Icn (Eq. 9) as follows:
q − 𝜎v /( pa )n 0.28
Qtn = t (7) m� = 1 −
pa 𝜎v� I
cn 25 (11)
1 + ( 2.65 )

− 0.15 ≤ 1.0
( 𝜎� )
n =0.381(Ic ) + 0.05 v
(8) OCR = 0.33 × Qtn (12)
pa 2.
by Kulhawy and Mayne (1990) and Robertson
√ (2009), where Qtn is defined in Eq. 7. This cor-
Icn = (3.47 − log Qtn )2 + (log Fr + 1.22)2 (9) relation is only valid for fine-grained soils.

where Fr is the normalized friction ratio, defined

5.2.3 Strength Parameters
as Fr = fs ∕(qt − 𝜎v ) 100%.
The parameters circumscribed by Eqs. (4)–(9) are
Similar to OCR, several methods are provided in the
internally calculated in module 2 (cf. Fig. 12) before
database to derive the effective friction angle (𝜑′). In
invoking module 4 (cf. Fig. 12). In other words, these
this study, the following method is selected:
parameters could be treated as source parameters
(similar to the CPT raw data) in the generated graphs. 𝜑� = 29.5◦ B0.121 [0.256 + 0.336Bq + log(Qt )] (13)
This also highlights the importance of selecting an
q

adequate method for assessing the unit weight as it by Mayne et al. (2009), where Qt and Bq are defined
influences the total and effective stress, which in turn in Eqs. 5 and 6, respectively. The valid range for this
influences the calculated “CPT parameters”. correlation is 0.1 ≤ Bq ≤ 1.0 and 20◦ ≤ 𝜙′ ≤ 45◦ and
it is only applicable for fine-grained soils.

Vol.: (0123456789)
13
 Geotech Geol Eng

5.2.4 Shear Wave Velocity by Mayne (2007), where m∗ = 0.6 for clean

quartz sands, 0.8 for silts and 1.0 for intact clays
The small-strain shear modulus (G0) is determined of low to medium sensitivity.
from in-situ Vs measurements. In case the shear wave The 1-D constrained tangent modulus M is gener-
velocity measurements are not available, correlations ally used to compute settlements. The following
relating the shear wave velocity to CPT results could correlation is used to determine the constrained
be used to estimate Vs. The methods database consists modulus:
of 15 methods to assess Vs, however, in this contribu-
tion the following four methods are selected:
M = 𝛼M (qt − 𝜎v ) (21)
Robertson (2009) suggested an approach based on Icn
Vs = 3.18q0.549 fs0.025 (14)
1.
c (Eq. 9) to determine 𝛼M:
by Hegazy and Mayne (1995); this correlation is
only valid for clays. • If Icn > 2.2:
 𝛼M = Qtn (if Qtn < 14)
Vs = (10.1 log qc − 11.4)1.67 (fs ∕qc × 100)0.3 (15)  𝛼M = 14 (if Qtn > 14)
2. • Otherwise if Icn < 2.2:
by Hegazy and Mayne (1995); this correlation is  𝛼M = 0.03[10(0.55Icn +1.68) ]
valid for different types of soils.
It has to be pointed out that compared to the methods
Vs = 1.75(qc )0.627 (16)
3. and parameters presented in Sects. 5.2.1, 5.2.2, 5.2.3
by Mayne and Rix (1995); this correlation is and 5.2.4, the current version of the database consists
only valid for clays. of numerous additional methods and parameters not
considered in this work. The main reason for select-
Vs = [𝛼vs (qt − 𝜎v )∕pa ]0.5 (17) ing a limited number of methods and parameters in
4.
this contribution is motivated by a reduction of com-
by Robertson (2015), where 𝛼vs = 10(0.55Icn +1.68);
plexity and increase of clarity; generating the graphs
this correlation is valid for different types of
with the current version of the database would lead
soils.
to overloaded graphs involving an unpractical number
nodes and edges that would, in turn, make the task
5.2.5 Stiffness Parameters of tracing different paths cumbersome to the readers.
Inversely, generating the graphs with a limited num-
The small strain shear modulus (G0) is deduced from ber of methods and parameters leads to comprehensi-
the shear wave velocity as follows: ble graphs (e.g, Fig. 14) that allow to trace back dif-
ferent paths. This is demonstrated by Fig. 15, which
G0 = 𝜌Vs2 (18) presents the output of module 4 in case the number of
where 𝜌 is the mass density of the soil, defined as methods and parameters is increased (e.g., 21 differ-
𝜌 = 𝛾t ∕g, in which g is the gravitational acceleration. ent values were computed for G0 located at the lower
Similar to the other parameters, a considerable num- left corner of the figure). As the main purpose of
ber of additional methods for computing G0 can be Sect. 5 is to present the graph-based approach imple-
selected from the database. Two of these methods are mented in APD in an illustrative way, simpler graphs
herein considered:

G0 = 2.78q1.335
c (19) Table 3  Machine learning based stratification of CPT 5/12
1.
by Mayne and Rix (1993); this correlation is Layer Start (m) End (m)
only valid for clays.
Layer 1—coarse-grained soil 0.00 8.50
G0 = 50pa (qt − 𝜎v )m∗
(20) Layer 2—transitional-soil 8.50 21.90
2. Layer 3—fine-grained-soil 21.90 40.00

Vol:. (1234567890)
13
Geotech Geol Eng

Fig. 14  Output of module 4 for selected methods

were generated by using a limited number of methods of these parameters were used as input parameters
and parameters. for the selected methods presented in Sects. 5.2.1,
5.2.2, 5.2.3 and 5.2.4. For each layer, the first and
5.3 Determination of Soil Parameters last measurements are filtered out. This procedure
reduces the effects of the neighbouring layers. For
This subsection demonstrates the output of the APD. layer 3 the averaged CPT data yield qc = 1392.8 kPa,
The two CSV files (imported by the system as inputs) fs = 29.4 kPa and u2 = 1008.7 kPa at 30.95 m depth.
include the parameters and methods presented in The groundwater level (GWL) is located at 3 m below
Sects. 5.2.1, 5.2.2, 5.2.3 and 5.2.4. the ground level. The cone tip net area ratio is pro-
The CPT data processed in this study is derived vided in the CPT GEF file as 0.85 (a = 0.85). The
from the test site described in Sects. 2, namely CPTu unit weight of water is defined as 10 kN∕m3. The
5/12 (Fig. 2). The CPTu soundings collected for this atmospheric pressure ( pa) corresponds to 100 kPa.
study revealed issues with the u2 measurements, par- Figure 14 shows the generated graph for layer 3.
ticularly in Layer 1 (coarse-grained material) and Two types of nodes are presented in the graph.
Layer 2 (sand-silt alteration). However, u2 measure- Parameters are represented by green nodes, while
ments were only used in Layer 3 (fine-grained mate- blue nodes correspond to methods. The links between
rial) to determine the friction angle (Eq. 13). Conse- parameters and methods within the system are illus-
quently, these issues did not affect the results of the trated by the arrows. These arrows have a defined
other sections. direction, either going from a parameter to a method
The ML-based stratification of this CPT is pre- or from a method to a parameter. The graph displays
sented in Sect. 3.5 and Fig. 5a. The first steps within the selected parameters and methods, while Table 4
APD were as follows: The system imported the CPT describes the parameters. Table 4 also illustrates the
GEF file, whereas the layers were provided manually values of the source parameters. These source param-
according to Table 3. For each layer, a SBT (accord- eters are either averaged CPT measurements or aver-
ing to Table 2) was assigned. Focusing on the 3rd aged initial parameters that were computed before
layer, SBT(3) (clay) was selected for this layer as generating the graph (e.g., total and effective stress).
estimated by the ML-based stratification. Know- In this case study, Eq. 4 was used to calculate the unit
ing the boundaries of the layer, the CPT raw data weight. As the system is created in an adaptable way,
(qc, fs and u2) as well as the parameters calculated the user could select other correlations for computing
from Eqs. (4)–(9) were averaged. The average values the initial unit weight. The output of each method is

Vol.: (0123456789)
13
 Geotech Geol Eng

Table 4  Parameters represented in graph (Fig. 14)

Parameters Notation used in the graph Description Comments/values

Source parameters Depth Middle of the layer 30.95 m

qc Cone tip resistance 1392.8 kPa
fs Sleeve friction 29.4 kPa
pa Atmospheric pressure 100 kPa
Icn Soil behaviour type index (Eq. 9) 3.41
SBT Soil behaviour type 3
Qtn Normalized cone resistance with vari- 3.36
able stress exponent (Eq. 7)
Bq Normalized porewater pressure (Eq. 6) 0.76
Qt Normalized cone resistance (Eq. 5) 3.35
gamma_sat_init Saturated unit weight obtained from 17.32 kN/m 3
Eq. 4
sigma_tot Total vertical stress 573.03 kPa
qt Corrected cone tip resistance 1544.13 kPa
sigma_eff Effective vertical stress 293.53 kPa
Intermediate parameters m3 Exponent m∗ used in Eq. 20 1
m2 Exponent m′ used in Eq. 10 and calcu- 1
lated from Eq. 11
alphaM Coefficient to determine constrained 3.36
modulus (Eq. 21)
sigma_p Preconsolidation stress determined 319.37 kPa
from Eq. 10
Vs Shear wave velocity 191.78, 187.26, 163.78, 184.14 m/sec
Destination parameters OCR Overconsolidation ratio 1.09, 1.11
phip Peak friction angle 29.68◦
G0 Shear modulus at small strain 64962.54, 61932.96, 47377.17,
59885.99, 48555.01, 45184.02 kPa
M_CPT Constrained modulus obtained from 3263.63 kPa
CPT

presented on the link connecting this method to the corner in Fig. 14, the selected method resulted in a
intermediate / destination parameter (e.g., output of friction angle of 29.68◦.
method_to_OCR_2 is 1.11). The value of intermedi- The constrained modulus was obtained from the
ate / destination parameter is shown next to the node selected method defined by Eq. 21, which resulted in a
(e.g., 2 values were computed for OCR, 1.09 and value of 3263.63 kPa. As a side note, the prediction of
1.11). the constrained modulus could be improved by using
It should be pointed out that it was not possible to other in-situ tests; to this end, the DMT (Marchetti
compare the obtained values (for different parameters) 1980) is currently incorporated in the APD frame-
to reference values, as no laboratory tests could be work. Generally, the incorporation of a higher number
adopted from the test site under study. Nonetheless, of in-situ tests to derive the parameters (and compare
the main aim of this case study is to present proof of the outcome of the different in-situ tests) leads to an
concept of the parameter determination framework. increase of confidence in the derived parameters.
Starting with the OCR located at the lower left corner The shear wave velocity (located at the right side
in Fig. 14, the selected two methods used to calculate in Fig. 14) is computed by four selected methods.
OCR resulted in values of 1.09 and 1.11, respectively. These methods resulted in values of 191.78, 187.26,
Moving to the friction angle located at the upper right 163.78 and 184.14 m/s, respectively. The shear

Vol:. (1234567890)
13
Geotech Geol Eng

modulus at small strains is located at the lower right to constitutive model parameters is considered as one
side in Fig. 14. Selected methods_to_G0_2 and 3 of the key aspects of the research project.
resulted in 48,555 and 45,184 kPa, respectively. As Currently, APD is able to interpret CPT and DMT
four different values of Vs were computed, method_ (Marzouk et al. 2023b) data; an extension of the
to_G0_1 results in four different values, 64962.54, framework (by including additional in-situ tests and
61932.96, 47377.17 and 59885.99 kPa. Ongoing measured in-situ shear wave velocity), validation
validation studies (not shown here) indicated that the of the output (by comparing the output of APD to
reliability of methods used to determine Vs is highly interpreted values from laboratory tests) and update
site dependent. Moreover, as could be expected, using and improvement of the database is part of ongoing
more accurate values of Vs will result in a more accu- research.
rate prediction of G0, which underlines the neces-
sity of creating a shear wave velocity module that is
able to import in-situ Vs measurements. This planned 6 Conclusions
module will be incorporated in module 1 (Fig. 12),
where in-situ Vs will be imported and used in further This work examines three different CPT data inter-
analyses. pretation strategies, with particular focus on the
identification of soil layer boundaries, spatial vari-
ability of CPT readings and soil parameter determi-
5.4 Final Comments Related to APD nation. Theoretical aspects and relevant use cases of
the underlying workflows, as well as relative merits
Figure 14 presents the generated graph for a clay and limitations compared to related approaches docu-
layer using a limited number of methods. In case mented in the literature are discussed throughout this
more methods are used (e.g., all of the methods in manuscript. The applicability and capabilities of the
the standard database of methods), this will lead to CPT data interpretation strategies are demonstrated
a considerable scatter in some of the derived param- considering in-situ CPT recordings recovered from
eters; cf. Fig. 15. Evaluating the scatter and selecting a test site in Salzburg, Austria. The primary objec-
a suitable approach for choosing a specific value from tive of this work is to advance towards an automated
the range of the calculated values is part of ongoing approach for layer detection, parameter determina-
research; the same applies to the process of validating tion, and evaluation of fluctuation degree, and to con-
and updating the correlations database. This is done nect it with numerical analysis.
by comparing the output of different correlations to In Sect. 3, it is demonstrated that the suggested
available laboratory tests; cf. Marzouk et al. (2023a). ML-based framework for the automatic identifica-
The uncertainties associated with using CPT data tion of soil layers predicts soil boundary distributions
for parameter determination are not currently taken that are well in agreement with the reference solution;
into account. In the current version of APD, the sys- hence, it offers a powerful alternative to classical filter
tem can only process one CPT at a time. The con- algorithms based on classification charts. Moreover,
sideration of several CPTs at the same time (e.g., the main motivation of automating the layer detec-
executed close to each other or from the same pro- tion process is to integrate it with other analyses, such
ject) would help to investigate whether the uncertain- as numerical analysis. Additionally, a ground model
ties due to the execution of the CPTs influence the approach adapted for "real-time" updates by modi-
obtained value(s) for the parameters or not. fying the soil stratification when new information
In this study, the transition from soil parameters to becomes available is being planned. Nevertheless, it
constitutive model parameters was not considered in is also highlighted that the accuracy of this approach
detail. Several correlations between soil parameters is considerably influenced by the quality and quantity
and some constitutive models, such as the Harden- of the database employed; in this respect, parameters
ing Soil model with small-strain stiffness (HSsmall) of the ML and KDE models should be updated and
(Benz 2007), are already available in the methods evaluated whenever the database is extended with
database. The transition from the CPT measurements additional samples. Furthermore, there is still a lack
of studies regarding interpolation strategies between

Vol.: (0123456789)
13
 Geotech Geol Eng

different soundings. Related work is currently under

consideration.
Section 4 showcases a data-driven strategy for
dynamically updating the CPT investigation layout.
Unlike conventional decision-making tools for gain-
ing maximum insight to the stratigraphic model at
minimum costs that mainly invoke Tobler’s First Law
of spatial analysis (Miller 2004) (i.e., all positions
are related but nearby positions are more related than
distant positions), the presented procedure involves
geotechnical sampling information as well. As a key
outcome, a scalar-valued field variable termed "level
of fluctuation" is calculated which allows for a sys-
tematic reduction of stratigraphic uncertainty in test
fields where CPT data is available. The efficacy is
demonstrated by means of both synthetic and real
case test sites, in combination with a comprehensive
discussion of potential application scenarios. The
latter include (but in not limited to) the employment
as intuitive tool in the communication of geo-spatial
uncertainty to project stakeholders or the objectifica-
tion of the CPT layout design. It is believed that the
level of fluctuation is a good starting point for any
sensitivity analysis. In the present work, this CPT
data interpretation strategy is solely deployed to cone
resistance recordings (single-factor approach). The
simultaneous consideration of multiple in-situ data
types (multi-factor approach), including the sleeve
friction, is beyond the capabilities of the current ver-
Fig. 15  Output of module 4 in case more methods and parameters are used

sion and will be addressed in future work.

Section 5 highlights a parameter determina-
tion framework (APD) that relies on a graph-based
approach to determine parameters based on in-situ
tests. This is extremely useful in the early stages of
projects when limited soil data is available. At this
stage, (relatively inexpensive) field tests such as CPT
and DMT are carried out prior to a full laboratory test
campaign. However, by using APD in the early design
phase of the project, users can efficiently obtain much
more detail. The motivation is not to replace labora-
tory tests with in-situ tests. They will still be needed
to improve the soil and constitutive model param-
eters for the final design. Furthermore, APD aims
to automatically connect determined parameters to
finite element (FE) software for numerical analy-
sis. The system has two key features, transparency
and adaptability. This means that users can add their
knowledge and experience to the system by extend-
ing the standard database of methods and parameters

Vol:. (1234567890)
13
Geotech Geol Eng

currently provided by the system. However, identify- Benz T (2007) Small-strain stiffness of soils and its numerical
ing the representative value for different parameters consequences. PhD thesis, University of Stuttgart
Brinkgreve RBJ (2019) Automated model and parameter selec-
is the biggest challenge due to the wide range of tion. Geostrata. https://​doi.​org/​10.​1061/​geosek.​00001​15
values obtained from the various correlations (meth- Bustamante M, Gianesselli L (1982) Pile bearing capacity pre-
ods) used. Therefore, a statistical module is currently diction by means of static penetrometer CPT. In: Balkema
being developed to aid in selecting the representative AA (ed) Proceedings of the 2nd European symposium on
penetration testing, Amsterdam, pp 493–500
value. Moreover, APD is currently undergoing expan- Chakraborty A, Dutta T, Mondal S, Nath A (2018) Applica-
sion and validation. tion of graph theory in social media. Int J Comput Sci Eng
6:722–729. https://​doi.​org/​10.​26438/​ijcse/​v6i10.​722729
Acknowledgements The authors would like to acknowl- Ebid AM (2021) 35 years of (AI) in geotechnical engineering:
edge the members of APD group, Arny Lengkeek and Ronald state of the art. Geotech Geol Eng 39(2):637–690. https://​
Brinkgreve. doi.​org/​10.​1007/​s10706-​020-​01536-7
EN ISO 14688-1 (2019) Geotechnical investigation and test-
Funding Open access funding provided by Graz University ing—Identification and classification of soil: part 1: iden-
of Technology. No funding was received to assist with the tification and description
preparation of this manuscript. Fattahi H, Babanouri N (2017) Applying optimized support
vector regression models for prediction of tunnel boring
Availability of Data and Materials The database used in machine performance. Geotech Geol Eng 35(5):2205–
this study is available at https://​www.​tugraz.​at/​insti​tute/​ibg/​ 2217. https://​doi.​org/​10.​1007/​s10706-​017-​0238-4
forsc​hung/​numer​ische-​geote​chnik/​daten​bank/ Felic H, Schlicke D, Granitzer AN, Tschuchnigg F (2023)
Enhanced interoperability between geotechnical and
Declarations structural engineering for 3D building models. RILEM
Bookseries, vol 43. Springer, pp 217–228. https://​doi.​org/​
Conflict of interest The authors declare no conflicting inter- 10.​1007/​978-3-​031-​33211-1_​20
ests. Giangiulio M, Granitzer AN, Tschuchnigg F, Hoffmann J
(2022) BIM-to-FEM: development of a software tool to
increase the operational efficiency of dam construction
Open Access This article is licensed under a Creative Com-
projects. https://​doi.​org/​10.​1007/​978-3-​031-​20241-4_​13
mons Attribution 4.0 International License, which permits
Goldsworthy GS, Jaksa MB, Kaggwa WS, Fenton GA, Grif-
use, sharing, adaptation, distribution and reproduction in any
fiths DV, Poulos HG (2004) Cost of foundation failures
medium or format, as long as you give appropriate credit to the
due to limited site investigations. In: Wang CM, Murugap-
original author(s) and the source, provide a link to the Crea-
pan K (eds) Proceedings of the international conference
tive Commons licence, and indicate if changes were made. The
on structural and foundation failures. Singapore
images or other third party material in this article are included
Gomes Correia A, Cortez P, Tinoco J, Marques R (2013) Arti-
in the article’s Creative Commons licence, unless indicated
ficial intelligence applications in transportation geotech-
otherwise in a credit line to the material. If material is not
nics. Geotech Geol Eng 31(3):861–879. https://​doi.​org/​10.​
included in the article’s Creative Commons licence and your
1007/​s10706-​012-​9585-3
intended use is not permitted by statutory regulation or exceeds
Graham J, Shields DH (1985) Influence of geology and geo-
the permitted use, you will need to obtain permission directly
logical processes on the geotechnical properties of a plas-
from the copyright holder. To view a copy of this licence, visit
tic clay. Eng Geol 22(2):109–126. https://​doi.​org/​10.​1016/​
http://creativecommons.org/licenses/by/4.0/.
0013-​7952(85)​90042-0
Hamza O, Bellis A (2016) Investigation into the effect of
uncertainty of CPT-based soil type estimation on the
References accuracy of CPT-based pile bearing capacity analysis.
In: Icelandic geotechnical society, Proceeding of the
17th Nordic geotechnical meeting. https://​doi.​org/​10.​
Abu-Farsakh MY, Shoaib MM (2024) Machine learning mod-
13140/​RG.2.​2.​35614.​74565
els to evaluate the load-settlement behavior of piles from
Hegazy YA, Mayne P (1995) Statistical correlations between
cone penetration test data. Geotech Geol Eng. https://​doi.​
Vs and CPT data for different soil types. In: Proceedings
org/​10.​1007/​s10706-​023-​02737-6
of symposium on cone penetration testing (CPT’95), vol
Al-Sammarraie D, Kreiter S, Kluger MO, Mörz T (2022) Reli-
2, pp 173–178
ability of CPT measurements in sand—influence of spac-
Henke S, Lerch H (2020) Building information modeling
ing. Géotechnique 72(1):48–60. https://​doi.​org/​10.​1680/​
(BIM) in der Geotechnik—Aktuelle Entwicklungen und
jgeot.​19.P.​247
Erfahrungen. Bauingenieur 95(12):490–500
Anagnostopoulos A, Koukis G, Sabatakakis N, Tsiambaos G
Henke S, Ebers-Ernst J, Rust M, Schäferhoff G (2020) DAS
(2003) Empirical correlations of soil parameters based
BIM-Baugrundmodell im Kontext des BIM-Gesamt-
on cone penetration tests (CPT) for Greek soils. Geo-
modells—estmodellvorstellung unter Einbeziehung
tech Geol Eng 21(4):377–387. https://​doi.​org/​10.​1023/B:​
zahlreicher Fachgewerke. In: Bundesanstalt für Was-
GEGE.​00000​06064.​47819.​1a
serbau (ed) Digitalisierung in der Geotechnik—Von

Vol.: (0123456789)
13
 Geotech Geol Eng

der Entwicklung zur Anwendung eines digitalen Marzouk I, Tschuchnigg F, Brinkgreve RBJ (2023b) Expansion
Baugrundmodells of an automated system for determining soil parameters
Jaksa MB, Goldsworthy JS, Fenton GA, Kaggwa WS, Grif- using in-situ tests. In: Proceedings of the 10th European
fiths DV, Kuo YL, Poulos HG (2005) Discussion: towards conference on numerical methods in geotechnical engi-
reliable and effective site investigations. Géotechnique neering (NUMGE 2023), London. https://​doi.​org/​10.​
55(8):625–626. https://​doi.​org/​10.​1680/​geot.​2005.​55.8.​ 53243/​NUMGE​2023-​70
625 Mayne P (2007) In-situ test calibrations for evaluating soil
Jaksa MB, Goldsworthy JS, Fenton GA, Kaggwa WS, Griffiths parameters. In: Tan TS (ed) Characterisation and engi-
DV, Kuo YL, Poulos HG (2005) Towards reliable and neering properties of natural soils. Balkema, Lisse and
effective site investigations. Géotechnique 55(2):109–121. Exton. https://​doi.​org/​10.​1201/​NOE04​15426​916.​ch2
https://​doi.​org/​10.​1680/​geot.​55.2.​109.​59531 Mayne P (2014) Interpretation of geotechnical parameters from
Jaky J (1944) The coefficient of earth pressure at rest. In: Hun- seismic piezocone tests. In: Proceedings, 3rd international
garian (a nyugalmi nyomas tenyezoje). J Soc Hung Eng symposium on cone penetration testing (CPT14, Las
Arch (Magyar Mernok es Epitesz-Egylet Kozlonye) pp Vegas), vol 102, pp 47–73
355–358 Mayne P (2017) Stress history of soils from cone penetration
Kahlström M, Mortensen PA, Hauser C, Hansen Børner tests. Soils Rocks 40:203–216. https://​doi.​org/​10.​28927/​
N (2021) Use of a 3D stratigraphic model as tool for SR.​403203
improved communication and risk assessment in large Mayne P, Rix G (1993) Gmax-qc relationships for clays. Geo-
infrastructure projects. In: IOP conference series: earth tech Test J. https://​doi.​org/​10.​1520/​GTJ10​267J
and environmental science, vol 710, no 1, p 012038. Mayne P, Rix G (1995) Correlations between shear wave veloc-
https://​doi.​org/​10.​1088/​1755-​1315/​710/1/​012038 ity and cone tip resistance in natural clays. Soils Found
Krige DB (1951) A statistical approach to some basic mine 35(2):107–110
valuation problems on the Witwatersrand. J Chem Met- Mayne P, Coop M, Springman S., Huang AB, Zornberg J
all Min Soc South Afr. https://​journ​als.​co.​za/​doi/​pdf/​10.​ (2009) State-of-the-art paper (SOA-1): geomaterial behav-
10520/​AJA00​38223X_​4792 ior and testing. In: Proceedings of the 17th international
Kulatilake PHSW, Um JG (2003) Spatial variation of cone tip conference on soil mechanics and geotechnical engineer-
resistance for the clay site at Texas A &M University. ing, Alexandria, vol 4
Geotech Geol Eng 21(2):149–165. https://​doi.​org/​10.​ Mazzoccola DF, Millar DL, Hudson JA (1997) Information,
1023/A:​10235​26614​301 uncertainty and decision making in site investigation for
Kulhawy FH, Mayne P (1990) Manual on estimating soil prop- rock engineering. Geotech Geol Eng 15(2):145–180.
erties for foundation design. Electric Power Research Inst https://​doi.​org/​10.​1023/A:​10184​99222​495
Kumar D, Samui P, Kim D, Singh A (2021) A novel methodol- Miller HJ (2004) Tobler’s first law and spatial analysis. Ann
ogy to classify soil liquefaction using deep learning. Geo- Assoc Am Geogr 94(2):284–289
tech Geol Eng 39(2):1049–1058. https://​doi.​org/​10.​1007/​ Niazi F (2022) CPT-based geotechnical design manual, Vol-
s10706-​020-​01544-7 ume 1: CPT interpretation-estimation of soil properties.
Likaj R, Shala A, Mehmetaj M, Hyseni P, Bajrami X (2013) Purdue University. https://​doi.​org/​10.​5703/​12882​84317​
Application of graph theory to find optimal paths for the 346
transportation problem. IFAC Proc Vol 46(8):235–240. Oberhollenzer S, Marte R, Gasser D, Premstaller M, Leitich A
https:// ​ d oi. ​ o rg/ ​ 1 0.​ 3 182/​ 2 0130​ 6 06-3-​ X K-​ 4 037.​ 0 0031. (2019) Microstructure of the Salzburger Seeton—charac-
(15th IFAC Workshop on International Stability, Tech‑ terization based on cone penetration tests. Geomech Tunn
nology, and Culture) 12(4):340–351. https://​doi.​org/​10.​1002/​geot.​20190​0012
Liu J, Liu J, Li Z, Hou X, Dai G (2021) Estimating CPT Oberhollenzer S, Fankhauser A, Marte R, Tschuchnigg F,
parameters at unsampled locations based on kriging inter- Premstaller M (2021a) Characterization of microstructure
polation method. Appl Sci 11(23):11,264. https://​doi.​org/​ in silt-dominated sediments. In: Proceedings of the 6th
10.​3390/​app11​23112​64 international conference on geotechnical and geophysical
Lunne T, Robertson PK, Powell J (1997) Cone penetration site characterization
testing in geotechnical practice. Soil Mech Found Eng. Oberhollenzer S, Premstaller M, Marte R, Tschuchnigg F,
https://​doi.​org/​10.​1007/​s11204-​010-​9072-x Erharter GH, Marcher T (2021) Cone penetration test
Marchetti S (1980) In situ tests by flat dilatometer. J Geo- dataset Premstaller Geotechnik. Data Brief 34(106):618.
tech Eng Div 106(3):299–321. https://​doi.​org/​10.​1061/​ https://​doi.​org/​10.​1016/j.​dib.​2020.​106618
AJGEB6.​00009​34 ÖNORM B 1997-2 (2017) Eurocode 7: Entwurf, Berechnung
Marzouk I, Tschuchnigg F, Paduli F, Lengkeek HJ, Brinkgreve und Bemessung in der Geotechnik—Teil 2: Erkundung
RBJ (2022) Determination of fine-grained soil param- und Untersuchung des Baugrunds
eters using an automated system. In: Proceedings of the Palla N, Gudavalli S, Subedi B, Jao M (2008) Comparison
5th international symposium on cone penetration testing, among SPT, CPT and Texas cone penetration test. In:
Bologna. https://​doi.​org/​10.​1201/​97810​03308​829-​77 ASCE (ed) Proceedings of ASCE Texas section spring
Marzouk I, Oberhollenzer S, Tschuchnigg F (2023a) An auto- meeting, pp 1–10
mated system for determining soil parameters: case study. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B,
In: 8th international symposium on deformation charac- Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg
teristics of geomaterials, Porto V et al (2011) Scikit-learn: machine learning in Python. J
Mach Learn Res 12:2825–2830

Vol:. (1234567890)
13
Geotech Geol Eng

Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Samui P (2012) Application of relevance vector machine for
Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg prediction of ultimate capacity of driven piles in cohesion-
V et al (2021) Scikit-learn: user guide. https://​scikit-​learn.​ less soils. Geotech Geol Eng 30(5):1261–1270. https://​
org/​stable/​user_​guide.​html doi.​org/​10.​1007/​s10706-​012-​9539-9
Phoon KK, Zhang W (2023) Future of machine learning in Schnaid F (2009) In situ testing in geomechanics: the main
geotechnics. Georisk Assess Manag Risk Eng Syst Geo- tests. Taylor & Francis, London
hazards 17(1):7–22. https://​doi.​org/​10.​1080/​17499​518.​ Shu-Xi W (2012) The improved Dijkstra’s shortest path algo-
2022.​20878​84 rithm and its application. Procedia Eng 29:1186–1190.
Rana H, Pandit B, Sivakumar Babu GL (2023) Estimation of https://​doi.​org/​10.​1016/j.​proeng.​2012.​01.​110
uncertainties in soil using MCMC simulation and effect of Umar SK, Samui P, Kumari S (2018) Deterministic and proba-
model uncertainty. Geotech Geol Eng 41(8):4415–4429. bilistic analysis of liquefaction for different regions in
https://​doi.​org/​10.​1007/​s10706-​023-​02523-4 Bihar. Geotech Geol Eng 36(5):3311–3321. https://​doi.​
Rauter S, Tschuchnigg F (2021) CPT data interpretation org/​10.​1007/​s10706-​018-​0498-7
employing different machine learning techniques. Geo- Van Berkom IE, Brinkgreve RBJ, Lengkeek HJ, De Jong AK
sciences 11(7):265. https://​doi.​org/​10.​3390/​geosc​ience​ (2022) An automated system to determine constitutive
s1107​0265 model parameters from in situ tests. In: Proceedings of the
Rauter S, Tschuchnigg F (2022) Identification of soil strata 20th international conference on soil mechanics and geo-
from in-situ test data using machine learning. pp 37–44. technical engineering, Sydney 2022
https://​doi.​org/​10.​1007/​978-3-​031-​12851-6_5 Vanmarcke EH (1977) Probabilistic modeling of soil profiles. J
Ricceri G, Simonini P, Cola S (2002) Applicability of piezo- Geotech Eng Div 103(11):1227–1246. https://​doi.​org/​10.​
cone and dilatometer to characterize the soils of the Ven- 1061/​AJGEB6.​00005​17
ice Lagoon. Geotech Geol Eng 20(2):89–121. https://​doi.​ Weil J (2020) Digitale Baugrundmodelle im Tunnelbau—Sta-
org/​10.​1023/A:​10150​43911​091 tus, Chancen und Risiken. Geomech Tunn 13(2):221–236.
Robertson PK (2009) Interpretation of cone penetration tests— https://​doi.​org/​10.​1002/​geot.​20190​0078
a unified approach. Can Geotech J 46(11):1337–1355. Xie J, Huang J, Lu J, Burton G, Zeng C, Wang Y (2022) Devel-
https://​doi.​org/​10.​1139/​T09-​065 opment of two-dimensional ground models by combining
Robertson PK (2010) Soil behaviour type from the CPT: an geotechnical and geophysical data. Eng Geol. https://​doi.​
update. In: 2nd international symposium on cone penetra- org/​10.​1016/j.​enggeo.​2022.​106579
tion testing, Huntington Beach
Robertson PK (2015) Guide to cone penetration testing for geo- Publisher’s Note Springer Nature remains neutral with regard
technical engineering. In: Proceedings 3rd international to jurisdictional claims in published maps and institutional
symposium on cone penetration testing (CPT14, Las affiliations.
Vegas)
Robertson PK (2016) Cone penetration test (CPT)-based soil
behaviour type (SBT) classification system—an update.
Can Geotech J 53(12):1910–1927. https://​doi.​org/​10.​
1139/​cgj-​2016-​0044
Robertson PK, Cabal KL (2010) Estimating soil unit weight
from CPT. In: 2nd international symposium on cone pen-
etration testing, Huntington Beach

Vol.: (0123456789)
13

View publication stats