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

net/publication/344879720

Rigorous test of the performance of shear-wave velocity correlations derived

from CPT soundings: A case study for Groningen, the Netherlands

Article in Soil Dynamics and Earthquake Engineering · January 2021

DOI: 10.1016/j.soildyn.2020.106471

CITATIONS READS

12 323

6 authors, including:

Pauline Kruiver Ger de Lange

Koninklijk Nederlands Meteorologisch Instituut Deltares
53 PUBLICATIONS 2,065 CITATIONS 29 PUBLICATIONS 622 CITATIONS

SEE PROFILE SEE PROFILE

Mandy Korff Jan Folkert van Elk

Deltares Shell Global
49 PUBLICATIONS 532 CITATIONS 51 PUBLICATIONS 1,425 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Pauline Kruiver on 26 October 2020.

The user has requested enhancement of the downloaded file.

 Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Contents lists available at ScienceDirect

Soil Dynamics and Earthquake Engineering

journal homepage: http://www.elsevier.com/locate/soildyn

Rigorous test of the performance of shear-wave velocity correlations

derived from CPT soundings: A case study for Groningen, the Netherlands
Pauline P. Kruiver a, c, *, Ger de Lange a, Fred Kloosterman a, Mandy Korff a, Jan van Elk b,
Dirk Doornhof b
a
Deltares, Daltonlaan 600, 3584 BK Utrecht, the Netherlands
b
Nederlandse Aardolie Maatschappij B.V, Schepersmaat 2, 9405 TA Assen, the Netherlands
c
Royal Netherlands Meteorological Insititute KNMI, Utrechtseweg 297, 3731 GA De Bilt, the Netherlands

A R T I C L E I N F O A B S T R A C T

Keywords: An improved correlation between shear-wave velocity VS and cone penetration test (CPT) parameters was
Shear-wave velocity derived for the Groningen region, the Netherlands, using a set of 154 pairs of CPTs and seismic CPTs. An analysis
Soil properties of a high-quality subset showed that the subdivision into Holocene versus Pleistocene intervals or into cohesive
Cone penetration test
versus non-cohesive intervals did not significantly improve correlation, nor did filtering out of the transition
Empirical correlation
intervals between clay and sand in the non-cohesive dataset. In addition, using qt or qc resulted in similar
regression results. We therefore propose a single new equation and we applied this to all available CPTs in the
region. The analysis of the fitting bias showed that there is a trend relative to the measured VS. We therefore
recommend performing SCPT measurements when very soft soils (VS < ~100 m/s) are encountered, rather than
relying on the empirical correlation based on CPT parameters.
The comparison of the VS30 values from CPTs to the model VS30 of Groningen shows that correlation is
excellent: 90% of the calculated VS30 values from the CPTs fall within one standard deviation of the VS30 of the
corresponding geological zone of the model. Although the resulting CPT-VS correlation is specific for the Gro­
ningen region, the approach can be applied to other regions with a paired SCPT-CPT dataset.

1. Introduction derived for specific soil types or apply to all soils. Examples of soil type
methods for VS are given in Ref. [2] for cohesionless soils (sand) and in
There are numerous empirical correlations in literature which link Ref. [3] for cohesive soils (clays). Examples of VS for all soil methods can
Cone Penetration Test (CPT) derived parameters to the shear-wave ve­ be found in Ref. [1,4,5,7]. Robertson [1] stressed that empirical corre­
locity VS (e.g. Ref. [1–7]). CPTs are routinely acquired to investigate soil lations from CPTs should be fitted to the local or regional conditions. An
conditions on building sites, especially in the Netherlands and in other example of a regional correlation between VS and CPTs for Christchurch,
soft soil regions. Usually, the maximum depth of a CPT is around 30 m New Zealand, is given in Ref. [10]. Following [7] on the functional form,
below the surface. Conform American building codes [8] and Eurocode Arup was the first to derive a CPT-VS correlation for the Groningen re­
8 [9] or national guidelines, information on VS is required to assess gion, based on multi-regression of 64 SCPTs [11]. Since then, many
seismic site classification using VS30 (VS time-averaged over the top 30 more SCPTs have become available that facilitate improving the
m). In addition, VS profiles versus depth are required for site response empirical CPT-VS correlation for Groningen. Vasileiadis’s correlation
analyses. [11] and the other correlations from literature, however, do not
Common parameters featuring in the empirical VS correlations and distinguish between Holocene and Pleistocene sediments. This distinc­
measured by or derived from CPTs are cone tip resistance qc, cone tip tion might be important for the regional Groningen conditions.
resistance corrected for penetration pore pressure qt, sleeve friction fs, The intervals representing Holocene sediments and Pleistocene
soil behaviour type index Ic, soil density ρ, depth D, vertical effective sediments are usually clearly recognisable in a CPT sounding as the
stress σ′ v and ageing scaling factor ASF. Empirical correlations can be transition between them is often represented by a jump in qc and fs and a

* Corresponding author. Royal Netherlands Meteorological Insititute KNMI, Utrechtseweg 297, 3731 GA De Bilt, the Netherlands.
E-mail addresses: [email protected] (P.P. Kruiver), [email protected] (G. de Lange), [email protected] (F. Kloosterman), mandy.korff@
deltares.nl (M. Korff), [email protected] (J. van Elk), [email protected] (D. Doornhof).

https://doi.org/10.1016/j.soildyn.2020.106471
Received 21 January 2020; Received in revised form 31 July 2020; Accepted 25 September 2020
Available online 22 October 2020
0267-7261/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

correlation between cone parameters and VS in an SCPT provides a

proxy for the ageing scaling factor that is used in site response calcu­
lations. It is therefore reasonable to investigate the correlation between
CPT and VS for the two stratigraphic groups of sediments that can be
readily recognised in both CPTs and SCPTs. Another subdivision of the
sediments is between cohesive sediments (clay and peat) and
non-cohesive sediments (sands of various grain sizes).
The paper first describes the CPT and SCPT datasets that were used in
the analysis and the derivation of the empirical correlation (section 2).
Section 3 describes the method, consisting of the functional form of
choice and the different subdivisions of the datasets. The justification of
using the CPT parameters qt and fs, along with σ ′ v in the empirical cor­
relation to determine VS is provided by the Exploratory Factor Analysis
(EFA) and Principal Component Analysis (PCA) in section 4. Testing the
effect of different splits of the data is described in section 5. The final
empirical equation for VS from CPTs is presented in section 6 and dis­
cussed in section 7. This equation is then applied to the large CPT dataset
available in the Groningen region and compared to the available VS
model (section 8). The paper’s conclusions and recommendations are
included in section 9.

2. Datasets

There is a large dataset of CPTs in the Groningen region, consisting of

thousands of CPTs. In general, the older CPT soundings contain data on
fs and qc only, while the newer CPTs include fs, qc and penetration pore
pressure, and therefore qt can be derived. For standard CPTs in the
Netherlands, readings are made at 2 cm depth intervals. The maximum
CPT depth is usually 30 m, unless the sounding was halted at a shallower
depth upon reaching the maximum capacity of the truck or because the
target (e.g. Pleistocene sand) was reached. Several studies assessing the
earthquake hazard from induced earthquakes as a result from the
extraction of gas from the onshore Groningen gas field used a dataset of
~5700 CPTs from the national DINO database supplemented by a subset
from CPTs gathered by engineering companies. In recent years, many
more CPTs were gathered to assess the necessary strengthening of
buildings in Groningen. The CPT dataset is described in Ref. [13]. The
total number of CPTs in the region available for the present study is
6818, of which 2912 have a maximum depth between 20 and 30 m and
1515 have a depth larger than 30 m. The median distance between a CPT
and its nearest neighbour in the dataset is 43 m (25th-75th percentile
range is 19–99 m). The distributions of the nearest distances are not
Gaussian.
During recent years, the SCPT dataset in Groningen has also steadily
increased in size. However, the quality of the SCPT data varies. In most
of the older SCPTs (before 2014), a standard VS sampling interval of 1.0
m was used. Stratigraphy, however, generally does not correspond to the
1.0 m sampling interval. As a result, the measured VS values are
Fig. 1. Sampling intervals for an SCPT for an example column of clay, peat and “contaminated” by mixing different stratigraphies and lithologies. The
sand. (left), standard sampling intervals of 1.0 m (centre) and alternative effect of sampling intervals is shown in Fig. 1, for an example soil col­
sampling intervals (right). umn of 1.8 m clay on top of 0.5 m of peat on top of sand. Using a
standard sampling interval of 1.0 m results in a representative VS for clay
for the first meter, a mixed clay/peat VS for the second meter, a mixed
jump in VS in a SCPT. Apart from the confining pressure, VS is influenced peat/sand VS of the third m, etc. In these mixed cases, the VS values are
by factors such as void ratio, geologic age, friction and cementation, generally not representative for the soil type. A high-quality dataset of
overconsolidation ratio (OCR) and historic cyclic loading. These factors 27 SCPTs was acquired in 2015 to characterise the stations of the
can have different effects on Holocene or Pleistocene sediments. For accelerometer monitoring network in Groningen [14]. The sampling
example, the OCR for Peelo clays of Pleistocene age varies between intervals for VS were chosen to coincide with stratigraphy boundaries
approximately 4 and 6 in the Groningen region, while the OCR of (right column of Fig. 1) and resulted in a high quality and
Naaldwijk clays of Holocene age is typically around 2. The empirical litho-stratigraphically unbiased VS dataset. Processing of the SCPT data
correlation between cone tip resistance and OCR [12] generally shows a has also been improved. Several additional SCPT datasets have become
high degree of scatter, indicating that the variation in OCR between the available during the last years, e.g. along the Eemskanaal levee in the
Holocene and Pleistocene clays might not be fully expressed by the Netherlands. We have assembled all SCPT data from different cam­
variation in cone tip resistance. A comparable cone tip resistance for paigns, projects and sources into one database. Raw seismic data, if
Holocene and Pleistocene clay might therefore correspond to a different available, were screened for quality. VS values of poor-quality raw data
VS value. Several authors, (e.g. Ref. [7]), suggest that the difference in were removed from the dataset. All SCPTs with a co-located CPT (either

2
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 2. Left: location of SCPTs and CPTs in Groningen. Right: outline of the Groningen gas field in the Netherlands.

at the same location or within a few meters) have been included. The where VS is the shear-wave velocity (in m/s), qt is the corrected cone tip
median distance between an SCPT and its nearest neighbour in the resistance (in MPa), fs is the sleeve resistance (in MPa) and σ ′ v is the
dataset is 329 m (25th-75th percentile range is 97–983 m). There are vertical effective stress (in MPa). We aim for these empirical correlations
154 SCPT-CPT pairs in our combined dataset [13]. Each SCPT-CPT pair to be practically applicable, meaning that their implementation should
is co-located, i.e. the maximum distance between them is only a few be straightforward for anyone who has familiarity with performing
meters. The geographical distribution of the datasets is shown in Fig. 2. CPTs, without the need for further processing. In theory, it is possible to
The statistics of the VS values and CPT parameters are summarised in derive a correlation for each combination of stratigraphy and lithology
Table 1. The histograms are provided as Figure SM1 in the Supple­ as was done for the geologically based VS [15]. However, to use a
mentary material. The median VS for the cohesive sediments is 177 m/s, stratigraphy and lithology specific correlation, every newly measured
which is considered as soft soil (ground type D, [9]) or in some cases as CPT should first be classified and next the corresponding correlation
special soil (ground type S1) [9]. The median VS for non-cohesive sedi­ should be applied. For most geo-engineers, the distinction between e.g.
ments is 263 m/s (ground type C [9]). Holocene and Pleistocene sediments or between cohesive and
non-cohesive sediments is clear. A more detailed classification would
3. Methodology require far more knowledge and would limit the applicability of the new
correlations. Therefore, we examine the correlations for two relatively
There are many functional forms available for the empirical corre­ easy subdivisions of the data: one for Holocene versus Pleistocene in­
lation between CPT parameters and VS (e.g. Ref. [1–7]. The CPT method tervals or one for cohesive versus non-cohesive intervals in the (S)CPTs.
was originally designed as a tool for estimating bearing capacity. As In addition, we examine the difference between using qc or qt and
such, the CPT parameters represent proxies of soil behaviour and soil whether excluding non-cohesive intervals close to transitions with
stiffness. Seismic velocities are also used as proxies of soil stiffness and cohesive intervals (filtering) will improve the fitting quality of the
represent different aspects of the soil behaviour. We adopt the functional regression.
form of [7] that only uses the parameters measured by the cone and a In order to derive the new correlations, two steps were performed.
calculated approximation of the confining pressure. The general form is The first step is to prepare the datasets for analysis (Fig. 3) and the
given by: second step is to perform multi-regression analysis to derive the a, b, c,
VS = a⋅qt b ⋅fs c ⋅σ′v
d
(1) and d components in Eq. (1). First, the sampling intervals of SCPTs and
CPTs were aligned: CPTs generally have a reading every 2 cm, while the
SCPT has sample intervals varying between 0.3 m and 1.0 m. The VS
Table 1 values from the SCPTs were resampled at the CPT reading depths by
Statistics of VS and CPT parameters in the combined SCPT-CPT dataset. duplication. An alternative option is averaging the values for the CPT
parameters over the depth range of each VS layer. This option was
Statistics VS (m/s) qc (MPa) fs (MPa) σ′ v (MPa)
included during an earlier stage of the study and resulted in similar
25th percentile 148.1 1.26 0.03 0.05
outcomes. We opted for the duplication of VS values for each CPT
median 213.6 2.95 0.08 0.11
75th percentile 263.6 10.70 0.14 0.17 reading in the remainder of the analysis. Second, the datasets of (S)CPTs
average 208.6 6.85 0.09 0.12 were classified in terms of stratigraphy and lithology. This was done for
standard deviation 79.8 8.10 0.07 0.07 testing the different options of combining certain layers, e.g. all

3
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

the Groningen conditions (Fig. 3). Using correlations of borehole-CPT

pairs in the area, the soil types in the different nomogram fields are
classified. The final lithology result was cross-checked with the GeoTOP
information (0.5 m depth resolution) and the Ic-based classification
scheme published by Robertson and Cabal [18]. An example of a clas­
sified (S)CPT sounding is given in Fig. 4. The transition between Holo­
cene and Pleistocene sediments at 12.9 m depth can be easily
distinguished in the SCPT and in the CPT. The low VS value between 4.5
and 5.0 m depth corresponds to a soft peat layer.
The lithology at each depth was used to determine the bulk density
and derive the vertical effective stress at the level of the CPT reading
following methods given in the national standards [19]. Finally, two
splits of the (S)CPT datasets were performed: a split with one subset
containing Holocene data and one subset containing Pleistocene data
(based on stratigraphy) and a split with one subset containing cohesive
and one subset of non-cohesive sediments (based on lithology). In finely
layered sediments, the CPT readings are influenced by the layer above
the tip and sleeve [20,21]. We therefore also prepared a non-cohesive
subset where data intervals within the sand layers that were closer
than 20 cm from the transition to clay were excluded: the boundaries
were filtered from the dataset (Fig. 5).

4. Exploratory Factor Analysis and Principal Component

Analysis

In order to investigate whether the parameters qc, fs and σ ′ v suffi­

ciently explain the parameter VS, we have performed Exploratory Factor
Fig. 3. Classification nomogram of the training set following [17] with class Analysis (EFA) and Principal Component Analysis (PCA) on the full
boundaries adjusted for Groningen conditions. The colours indicate the density dataset. EFA is widely used in the social sciences and various statistical
of CPT data points. Field codes refer to: very dense fine/coarse sand (A1), dense software packages including tools to perform all sorts of EFA are avail­
fine sand (B1), moderately dense coarse sand (C1), loose coarse sand (D1), able. EFA and PCA have also found their ways to the earth sciences. Both
dense sand (A2), moderately dense silty sand (B2), moderately dense sand (C2), statistical techniques have been used frequently as numerical proced­
loose sand (D2), very stiff silty clay, silt (A3), stiff sandy clay (B3), moderately ures to screen variables, to trace linear dependencies of variables and to
dense clayey sand (C3), loose clayey sand (D3), very stiff clay (A4), stiff construct effective regression equations.
(organic) clay (B4), firm (organic) clay (C4), soft (organic) clay (D4), very
Although EFA and PCA look similar at first glance, there is a distinct
dense peat (A5), dense peat (B5), moderately dense peat (C5), unconsolidated
difference. EFA searches for linear combinations of variables, the so-
peat (D5). (For interpretation of the references to colour in this figure legend,
the reader is referred to the Web version of this article.)
called latent factors, which represent underlying fundamental quanti­
ties of processes manifested in the observed variables. PCA on the other
hand seeks linear combinations in the data matrix of assumed inde­
Holocene layers. Because of the relatively limited size of the SCPT-CPT pendent variables that are uncorrelated and of high variance. In other
dataset, this was performed in a semi-automated way. The stratigraphy words, it checks if it is possible to achieve data reduction in the CPT data
was based on the geological model GeoTOP from TNO Geological Survey matrix when the variables qc, fs and σ′ v show linear dependence. EFA
of the Netherlands at the (S)CPT location [16]. The stratigraphic tackles the complete dataset of VS, qc, fs and σ ′ v while PCA addresses the
boundaries were manually adjusted, because the depth discretisation of assumed independent variables qc, fs and σ′ v only.
0.5 m in the GeoTOP 3D grid generally cuts across the boundaries In the following analysis with EFA and PCA the full set of 154 SCPT/
observed in the (S)CPTs. The (S)CPTs represent local and more detailed CPT pairs is used without making any splits in the dataset. This is
information than the GeoTOP model. Therefore, the stratigraphic because various amounts of SCPTs became available at different stages
boundaries observed in the (S)CPT are prevailing in our interpretation. at the project. The latest available SCPTs were not included in the CPT-
The lithology at each CPT sounding was automatically determined using VS empirical equation testing different splits of the data, but they were
a classification based on Douglas and Olson [17]. We used an adaptation included in the EFA and PCA analysis.
of the original nomogram in which the friction ratio Rf is plotted on a
logarithmic scale (Fig. 4). This transforms the two groups of lithoclass 4.1. Exploratory Factor Analysis (EFA)
boundaries into straight lines and hyperbolas respectively approximated
by the functional forms The R foundation for Statistical Computing [22] offers a wide variety
a of statistical tools and advanced packages. In the R library several tools
qc = (2) are available to perform Factor analysis. For the processing of the CPT
Rf
data in this paper, the tool “fa” has been used, being part of the large
for the straight lines and package “psych” in R. Tools dealing with EFA offer several methods for
oblique transformations of the solution and likewise several factoring
200 b− 3
qc = ( ) (3) methods. In this case, “varimax” matrix rotation was applied, with
b − Rf
“maximum likelihood” as the factoring method, assuming multivariate
for the hyperbolas. Using the original four-fold subdivision in the normally distributed data. The Eigenvalues are determined initially,
nomogram results in 20 different possible fields to discriminate between using the “fa” tool, to estimate the number of underlying factors in the
soil groups. By varying the constants a and b the boundaries can be dataset VS, qc, fs and σ ′ v. Eigenvalues are plotted in so-called scree plots
shifted between clusters of qc-Rf to arrive at a site-specific nomogram for as shown in Fig. 6 (left). This plot shows the Eigenvalues for the original
data matrix and for the factor solution after varimax matrix rotation

4
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 4. Example of an SCPT from the

training set with code DKPS021. Left:
measured VS profile in the SCPT. The low VS
value between 4.5 and 5.0 m depth corre­
sponds to a soft peat layer. Middle: the cor­
responding CPT profile. The qc (blue) and qt
profiles (red dots) are very similar and plot
almost on top of each other. Right: the di­
vision in subsets from left to right: all (blue);
Holocene (orange) and Pleistocene (purple);
cohesive (green) and non-cohesive (yellow);
intervals filtered out (black) from the non-
cohesive dataset (yellow). Subset colours
correspond to Fig. 8. (For interpretation of
the references to colour in this figure legend,
the reader is referred to the Web version of
this article.)

with “fa” for increasing factor number. The point of levelling off (the approximately the same, 40% and 37%, respectively. Thus, although the
“elbow”) is found at factor number 3, indicating that 2 factors can be Eigenvalue of the second factor is much lower than the first one, it still
retained. The graphical representation of the retained factors is shown in accounts for a similar proportion of explained variance and hence
Fig. 6 (right). cannot be ignored.
The results of the Factor Analysis with “fa” are summarised in Both Table 2 and the graphical presentation in Fig. 6 (right) show
Table 2 as standardized factor loadings (matrix of weights). Table 2 one factor (1) related to qc and fs and the other factor (2) to VS and σ′ v.
shows that the proportion of explained variance for both factors is The factors can be interpreted in a geotechnical sense as the lithology

Fig. 5. Preparation of the datasets.

5
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 6. EPA results. Left: Scree plot of Eigenvalues of the data matrix VS, qc, fs and σ ′ v. Right: Graphical presentation of the retained 2 factors and their loadings (after
R library, “psych” package, [22]). Stress_eff refers to σ′ v, ML refers to Maximum Likelihood.

scree plot of Eigenvalues shows that the point of levelling off is found at
Table 2 factor 3. This indicates that at least 2 principal components can be
EPA - Factor loadings and Proportion of Explained Variance, significant loadings retained. The steepness of the scree plot between component number 2
in bold italics.
and 3 suggests that even 3 components might be retained.
Factor Loadings The results of the PCA analysis are summarised in Table 3. In case of
Variable Factor 1 (ML1) Factor 2 (ML2) assuming a model of 2 principal components (only qc and σ ′ v), 92% of
VS 0.55 0.68
the total variance can be explained, while a single component already
qc 0.85 0.33 reaches 76%. Retaining the maximum amount of 3 components results
fs 0.68 0.50 in single significant loadings (loadings > 0.7) per component and
σ′ v 0.32 0.82 practically equal proportions of explained variance. This supports the 3-
Proportion of Variance 0.40 0.37
component model of high variance for each of the components.
Cumulative Variance 0.40 0.77
In conclusion, the PCA analysis indicates that the variables qc, fs and
σ ′ v are statistically uncorrelated and responsible for equal variances in
component qc/fs in the first factor with an obvious large effect in the total CPT data matrix. It also implies that PCA supports the use of
observed VS and the second factor as gradual increase of VS with depth these 3 variables in the basic model to estimate VS.
and thus with increasing effective stresses with depth. Exchanging the
effective stress with depth as a variable in the factor analysis yields the 5. Testing different splits of the dataset
same results. We prefer using the effective stress, because of the pres­
ence of thick layers of low-density soil (i.e. peat). We tested the various splits of the dataset using a subset of 109
SCPTs. The high-quality data from the 27 SCPTs from Ref. [14] were
used as the training set, because they represent the highest quality. In
4.2. Principal Component Analysis (PCA) this dataset, SCPT VS intervals were chosen to correspond with stratig­
raphy boundaries (Fig. 1, right column). Two other separate lower
PCA is applied on the CPT data matrix qc, fs and σ ′ v, all expressed in quality datasets were used as test sets. Test set 1 comprises the SCPT data
MPa (n = 272,348) to test the hypothesis that these assumed indepen­ gathered up until 2014 (62 locations), generally with standard sampling
dent variables are uncorrelated and of high variance. The PCA tool intervals of 1.0 m or sometimes even 2.0 m. Test set 2 comprises of SCPT
“principal” in the package “psych” in R has been applied. Like EFA, a data from the Eemskanaal levee gathered in 2016 (20 locations, [23]).
varimax rotation has been used. The PCA tool generates the Eigenvalues
of a dataset to estimate the number of principal components (Fig. 7). The
Table 3
PCA - Standardized loadings and Proportion of Explained Variance, significant
loadings in bold italics.
Standardized loadings

1 2 Components 3 Components
Component

Variable Comp. 1 Comp. Comp. Comp. Comp. Comp.

1 2 1 2 3

qc 0.88 0.93 0.23 0.25 0.90 0.35

fs 0.91 0.81 0.44 0.33 0.39 0.86
′
σv 0.82 0.32 0.94 0.93 0.23 0.27
Proportion 76% 59% 41% 35% 34% 31%
Explained
Cumulative 76% 59% 92% 35% 69% 100%
Proportion

Fig. 7. PCA results, scree plot of Eigenvalues of the data matrix qc, fs and σ ′ v.

6
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 8. Options for splits of training data in the development phase.

Table 4
Numbers of options for splits of training data in the development phase and the resulting values for a, b, c and d in Eq. (1).
Number Filtered or unfiltered qc or qt Layers included a b c d

1 Unfiltered qc Holocene 186.6 0.204 − 0.048 0.169

2 Unfiltered qc Pleistocene 299.0 0.092 0.019 0.163
3 Unfiltered qc Cohesive 264.6 0.266 − 0.039 0.238
4 Unfiltered qc Non-cohesive 346.8 0.085 0.116 0.132
5 Unfiltered qc All 344.1 0.116 0.073 0.201
6 Unfiltered qt Holocene 200.1 0.187 − 0.010 0.144
7 Unfiltered qt Pleistocene 279.5 0.097 0.014 0.141
8 Unfiltered qt Cohesive 257.5 0.261 − 0.019 0.206
9 Unfiltered qt Non-cohesive 352.4 0.079 0.121 0.127
10 Unfiltered qt All 331.3 0.118 0.079 0.180
16 Filtered qc Holocene 184.8 0.219 − 0.055 0.171
17 Filtered qc Pleistocene 282.2 0.096 0.010 0.147
18 Filtered qc Cohesive 264.6 0.266 − 0.039 0.238
19 Filtered qc Non-cohesive 362.9 0.061 0.133 0.102
20 Filtered qc All 342.0 0.117 0.073 0.199
21 Filtered qt Holocene 198.8 0.201 − 0.015 0.144
22 Filtered qt Pleistocene 264.2 0.101 0.006 0.126
23 Filtered qt Cohesive 257.5 0.261 − 0.019 0.206
24 Filtered qt Non-cohesive 362.5 0.061 0.133 0.102
25 Filtered qt All 329.3 0.119 0.080 0.177

The sampling interval was 1.0 m. Later during the study, even more example, the Holocene fit on the Holocene dataset was compared to the
SCPTs became available. The additional SCPTs were not included in this “all” fit on the Holocene dataset.
part of the analysis, but they were included in the final empirical The measured VS and predicted VS profiles for the different options
equation. are plotted in Fig. 9 for DKPS011 and in Fig. 10 for DKPS021 from the
The options of the data splits are visualised in Fig. 8 and tabulated in training dataset. In general, the agreement between the predicted and
Table 4. All options were used in the regressions to derive the a, b, c, and measured VS is satisfactory. In some intervals, the predicted values are
d components of Eq. (1). This resulted in 20 sets of components, which higher than the measured values and in other intervals the opposite is
were applied to the corresponding parts of the training dataset and the observed. A consistent positive shift between the predicted and
two test datasets for the evaluation of the error metrics. In addition, measured VS over a certain interval can be present in a SCPT. An
several fits for “all” were applied to subsets for a cleaner comparison. For example is shown in the Pleistocene part (depth > 12 m) in DKPS011

7
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 9. Predicted VS (red) and measured VS (blue) for the different options for SCPT coded DPKS011 from the training dataset. The top row consists of unfiltered data,
the bottom row of filtered data. The options for the datasets are listed in Table 4. (For interpretation of the references to colour in this figure legend, the reader is
referred to the Web version of this article.)

(Fig. 9): the predicted VS is almost always higher below 15 m in this sand, because of the low permeability of clay and possible overpressures
particular CPT. However, in other SCPTs, this shift is not present (Fig. 9) in clay. The error metrics for the clay (cohesive, options 3 and 8),
or negative. however, are very similar. Therefore, we conclude qc and qt can both be
Several definitions were used to examine the goodness of fit. For each used in the empirical equations and that there is no need to discard the
datapoint, the bias is defined as the predicted value minus the measured CPTs where only qc was measured, as they are as valuable as the CPTu’s
value. The error metrics consist of the average negative bias, the average (i.e. CPTs with qt).
bias, the average positive bias and the root-mean-square error (RMS) for From Fig. 11 several other observations can be made. Focussing on qc
each subset. The predicted values were calculated only for the corre­ (left part of the figure), we observe that the average negative bias is
sponding dataset. For example, the Holocene fit was applied to the generally higher than the average positive bias. The number of samples
Holocene intervals of the training set and the two test sets. in the average negative bias is approximately equal to the number of
In the interest of practicality, we wanted to investigate whether it is samples in the positive bias. For example, for option 5 there are 9165
worthwhile to split the data into subsets or to use qt rather than qc. To datapoints with a negative bias in the training set and 9066 datapoints
this end, we analysed the error metrics for selected combinations of with a positive bias. Overall, the average bias is slightly negative for the
subsets. The first option is the difference between using qc or qt, shown in training dataset (− 4 m/s) and positive or negative for the test set 1
Fig. 11. In theory, qt provides more reliable results than qc. In older (range of average from − 6 to +13 m/s), depending on the data split
CPTs, however, the pore pressure has often not been measured and only option. The average bias for test set 2, however, is much larger (− 30 to
qc is available. In addition, pore pressure is not always measured in +27 m/s). This will be explained and corrected later. The error metrics
modern CPTs, because including the pore pressure measurement is of the test sets are similar to the training set for test set 1 or higher than
slightly more expensive. The first observation from Fig. 11 is that for the training set test set 2. This is to be expected, because the highest
each subset, the error metrics for qc and qt are very similar. As expected, quality data were used for training.
the effect of the correction for pore pressure is larger in clay than in Apart from using qt or qc, we investigated the effect of several data

8
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 10. Predicted VS (red) and measured VS (blue) for the different options for SCPT coded DKPS021 from the training dataset. The top row consists of unfiltered
data, the bottom row of filtered data. The options for the datasets are listed in Table 4. (For interpretation of the references to colour in this figure legend, the reader
is referred to the Web version of this article.)

splits. An obvious split is the distinction between Holocene and Pleis­ cohesive fit is worse. This might be related to the transition effect in the
tocene sediments. The error metrics for this split is shown in Fig. 12, top sand layers when the reading was close to a clay-sand transition. The
panel. The fit based on the subset of the data is compared with the fit error metrics for the comparison between unfiltered and filtered data is
based on the entire training set (“all”) applied to the subset. The error shown in Fig. 13. This figure shows that the effect of filtering is negli­
metrics for the fit based on all data and applied to all data is shown for gible: all error metric patterns look very similar for each dataset.
reference (right column, option 5). The error metrics are very similar for Therefore, we conclude that it is not worthwhile to apply the distinction
the Holocene (H fit compared with H fit all). This also holds for the between cohesive and non-cohesive intervals in the CPT data in order to
Pleistocene fit (P fit compared with P all fit). Comparing the splits to all convert to VS.
data: the Holocene performs slightly better, but the Pleistocene performs Fig. 11 showed that the second test set has very high error metrics
worse than the unsplit datasets. The overall conclusion is that it is not compared with the other two datasets. We applied the VS equation from
worthwhile to apply the Holocene/Pleistocene split in the CPT data in option 5 (no data splits, no filtering) to predict the VS values from the
order to convert to VS, because there is no significant improvement in CPT measurements for the three datasets and calculated the bias. The
accuracy. distribution of the bias is shown in Fig. 14. The top panel shows that the
A second subdivision is between cohesive (clay and peat) and non- bias is Gaussian and centred around 0 for the training set and test set 1.
cohesive (various grain size classes of sand) sediments. In general, the However, there is a shift in bias towards values larger than 0 for test set
Holocene sediments tend to be cohesive and the Pleistocene sediments 2. Moreover, there is high amount of bias values between − 150 and
are predominantly sandy. There are, however, also sandy Holocene and − 100 m/s. We have investigated the cause of these two deviations by
clayey Pleistocene sediments. Therefore, this split represents a different looking at the original SCPT data. We have reinterpreted the VS mea­
subdivision of the data than the Holocene/Pleistocene subsets. The error surements using the original seismic traces and found that a large part of
metrics for the cohesive/non-cohesive analysis are shown in Fig. 12, the bias could be attributed to the use of interval travel times measured
bottom panel. We observe a similar pattern as for the Holocene/Pleis­ with a dual sensor seismic cone [23]. After reinterpretation and cor­
tocene split. The distinction between cohesive and non-cohesive does recting the VS data from the SCPTs in test set 2, the distribution of this
not perform convincingly better than the unsplit datasets. Moreover, dataset is improved and centred around 0 (Fig. 14, bottom panel). We
although the cohesive fit is better than the “all” fit on all data, the non- have used the corrected VS data in the remainder of the analysis.

9
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 11. Error metrics for the three datasets using all unfiltered data, for qc (left half of the plots) and for qt (right half of the plots). The average negative bias (-av neg
bias) is expressed as a positive value for easy comparison with the other error metrics. Av pos bias = average positive bias, av bias = average bias, RMS = root-mean-
square error, H = Holocene, P = Pleistocene, coh = cohesive sediments. Non-coh = non-cohesive sediments.

6. New empirical equation for VS from CPTs 7. Discussion

So far, we have tested the performance of regression function ob­ Extensive analysis of the VS distributions and the bias (measured
tained from the high-quality training set. The conclusion from the error minus predicted VS) is provided in the Supplementary material and
metrics analysis is that it is not worthwhile to split the CPT data in either summarised here. The distributions of measured and predicted VS are
Holocene/Pleistocene intervals or in cohesive/non-cohesive intervals. shown in Fig. 16. The distributions have identical median and statistics
Moreover, qt and qc data can be treated alike. We therefore do not apply (Fig. 16, left). The tails of the distributions, however, deviate from a
any splits or filtering to the entire dataset of 154 SCPT-CPTs (test set 2 Gaussian distribution (Fig. 16, right).
data were corrected) before we repeat the regression analysis. We pre­ The bias for the input parameters of Eq. (4) is shown in Figure SM5 of
sent a single new empirical equation for the conversion of CPT param­ the Supplementary material. There is no systematic trend bias when
eters to VS, based on the full SCPT-CPT dataset: plotted versus qc, fs or σ′ v. The bias for measured and predicted VS is
shown in Fig. 17. The linear fit through the data points in Fig. 17 (left,
(4)
0.204
VS = 359.0⋅qc 0.119 ⋅fs 0.100 ⋅σ′v top panel) when considering predicted VS almost coincides with the 1:1
line, indicating that on average the fit is very good. This is supported by
where VS is the shear-wave velocity (in m/s), qc is the measured cone tip
an average bias (Fig. 17, left, bottom panel) that is close to 0. The bias
resistance (in MPa), fs is the sleeve resistance (in MPa) and σ′ v is the
shows characteristic oblique lines of points. This is caused by the fact
vertical effective stress (in MPa). Because qc and qt performed equally
that there are CPT readings – and thus predicted VS values – for one
well in the regression on the training set, Eq. (4) can also be applied to qt
value of measured VS from the corresponding SCPT. There is, however, a
instead of qc. Several examples of profiles of measured and predicted VS
trend for the measured VS (Fig. 17, right panels). The predicted VS is
using Eq. (4) are shown in Fig. 15. This figure shows that the predicted
underestimated for higher VS and overestimated for lower VS values. The
VS generally follows the measured trend, but values are lower in some
fact that there is a trend in the bias suggests that not all factors deter­
intervals and higher in other intervals.
mining the VS value are captured by the CPT parameters, since both are

10
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 12. Error metrics for the training set with qc and unfiltered data. Top: testing the split between Holocene and Pleistocene data. From left to right: the Holocene
fit on the Holocene subset, the “all” fit on the Holocene subset, the Pleistocene fit on the Pleistocene subset, the “all” fit on the Pleistocene subset and the “all” fit on
all data of the training set. Bottom: testing the split between cohesive and non-cohesive material. From left to right: the cohesive fit on the cohesive subset, the “all” fit
on the cohesive subset, the non-cohesive fit on the non-cohesive subset, the “all” fit on the non-cohesive subset and the “all” fit on all data of the training set. See
caption of Fig. 11 for abbreviations.

Fig. 13. Error metrics for the training set with qc for testing the filtering of non-cohesive layers. From left to right: the unfiltered non-cohesive fit on the unfiltered
non-cohesive subset, the unfiltered “all” fit on the non-cohesive unfiltered subset, the filtered non-cohesive fit on the filtered non-cohesive subset, the filtered “all” fit
on the non-cohesive filtered subset. See caption of Fig. 11 for abbreviations.

different proxies of soil behaviour. The underprediction of relatively 8. Application of the new empirical equation to the Groningen
high VS is generally not a problem in geotechnical applications, because CPT dataset
this leads to an overprediction of e.g. site effects. The overestimation of
the relatively low VS values, however, poses a problem. These lower VS The new empirical Eq. (4) has been applied to the Groningen CPT
values are representative of very soft soils, such as peat. One of the dataset [13], resulting in a VS profile for each CPT location. One way to
factors contributing to site amplification is the VS profile. For the Gro­ define a metric that can be compared to a different data source of VS is to
ningen region, low VS30 values generally correspond to higher amplifi­ convert the VS profiles into VS30, which is the time-averaged VS over the
cation factors [24]. The overprediction in VS generally leads to an top 30 m. This can only be done if the CPT is sufficiently long. Standard
underprediction of the amplification. Considering the trend in the bias, CPTs typically reach to 30 m depth, which is convenient for the calcu­
we therefore recommend performing SCPT measurements for these lation of VS30. Especially in the southern part of the region, however,
shallow, soft sediments, rather than relying on VS values predicted from CPTs do not always reach the full 30 m, because of the maximum load
a CPT combined with a correction factor for the bias. capacity of the truck. Strictly speaking, CPTs shorter than 30 m are not
suitable for VS30 calculation. In order to avoid discarding too many CPT
datapoints, we have included CPTs with a length between 29 and 30 m

11
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

for Groningen [15]. This paper describes the procedure of deriving the
VS30 map for the Groningen region and is summarised here. A subset of
SCPTs, available in 2017, was used to derive empirical correlations of VS
versus confining stress for all soil types present in the Groningen region.
The empirical correlations include average VS values and standard de­
viations. The soil columns from the detailed geological model GeoTOP
[16] were converted to VS profiles using the soil type of each layer and
the corresponding empirical correlation including the uncertainty. Each
VS profile was then converted into a VS30 value. The VS30 values were
aggregated into zones of similar geological built-up and described by the
average VS30 and standard deviation for each zone. The VS30 values from
the CPTs are plotted on the map of the VS30 model [15] in Fig. 18. The
VS30 values from the CPTs represent point observations, while the VS30
model represents a regional model. Overall, the agreement is excellent:
90% of the calculated VS30 from CPTs fall within the average plus or
minus one standard deviation of the modelled VS30 from the Ground
Motion Model. Apparently, the overprediction for relatively low VS and
the underprediction for relatively high VS observed in the Discussion
section are averaged out over the 30 m interval. An exception to good
agreement between CPT and model is to be found in the Eemshaven area
in the north. The VS30 from the CPTs is significantly higher than from the
GMM VS model. This is due to the fact that the CPTs are situated on
newly reclaimed and geotechnically improved land, while the VS model
is based on the GeoTOP model with average properties of the anthro­
pogenic soil type.

9. Conclusions and recommendations

We have combined the CPT and SCPT data from the Groningen (the
Netherlands) database to create an empirical correlation between the VS
and the CPT parameters cone tip resistance (qc or qt), sleeve resistance
(fs) and vertical effective stress (σ′ v). The applicability of the new
empirical correlation increases when no interpretation steps such as
classification of the data are required. In addition, either qc or to qt can
be used in the empirical equation. Before we presented the single new
equation for the region of Groningen, we have analysed the performance
of various regressions on different data splits. We showed that a subdi­
vision of the dataset into Holocene versus Pleistocene intervals or into
cohesive versus non-cohesive intervals did not significantly improve the
prediction of VS from CPTs. In addition, filtering the transition intervals
between clay and sand in the non-cohesive dataset also did not signifi­
cantly improve the prediction. Based on this analysis, we conclude that
it is valid to apply a single empirical equation to the entire CPT depth
range.
The analysis of the bias showed that there is a trend relative to the
measured VS. The relatively low VS values are overpredicted, leading to
an underestimation of e.g. amplification in site response. We therefore
recommend performing SCPT measurements when these very soft soils
(VS < ~100 m/s) are encountered, rather than relying on the empirical
correlation based on CPT parameters.
The new equation was applied to the Groningen CPT database,
Fig. 14. Distribution of the bias for the fit on all unfiltered data using qc (option
provided that the CPTs were at least 20 m deep. The predicted VS profiles
5) for the training set, test set 1 and original test set 2 (top) and for the training
were transformed to VS30 values. 90% of the calculated VS30 from CPTs
set, test set 1 and corrected test set 2 (bottom).
fall within the average plus or minus one standard deviation of the
modelled VS30 from the Ground Motion Model [15].
in the VS30 calculation, correcting for the missing maximum 1 m of
length. The approach results in 2103 datapoints of VS30 derived from
CPT parameters with an acceptable geographical spread of VS30 values
from CPTs (Fig. 18). The other source of VS information is the VS model

12
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 15. Predicted VS (red) and measured VS (blue) for several SCPTs using Eq. (4). (For interpretation of the references to colour in this figure legend, the reader is
referred to the Web version of this article.)

Fig. 16. Left: Boxplots with outliers, showing the mean (thick line in the box) the median (thin line in the box), 25th and 75th percentiles (limits of the box), 1.5x the
difference between the 25th and 75th percentile (whiskers) and outliers (circles). Right: Probability plot for predicted VS assuming a normal distribution (red line).
(For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

CRediT authorship contribution statement Resources, Funding acquisition. Dirk Doornhof: Resources, Funding
acquisition.
Pauline P. Kruiver: Methodology, Validation, Software, Writing -
original draft, Writing - review & editing, Visualization. Ger de Lange: Declaration of competing interest
Methodology, Formal analysis, Software, Writing - review & editing.
Fred Kloosterman: Methodology, Formal analysis, Software, Writing - The authors declare that they have no known competing financial
review & editing, Visualization. Mandy Korff: Conceptualization, interests or personal relationships that could have appeared to influence
Writing - review & editing, Project administration. Jan van Elk: the work reported in this paper.

13
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

Fig. 17. Predicted and measured VS for the entire dataset of 154 SCPT-CPT pairs. Top: VS data with 1:1 gradient (blue) and the linear regression line (green). Bottom:
bias for entire dataset. The green line represents the linear fit through the bias. Left: plots versus predicted VS; right: plots versus measured VS. (For interpretation of
the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Acknowledgements

We would like to thank Renée de Bruijn, Jan Gunnink and Jan Stafleu
from TNO Geological Survey of the Netherlands for their discussion
about CPT and VS during an early stage of the analysis. This work was
funded by Nederlandse Aardolie Maatschappij B.V.. Constructive com­
ments by an anonymous reviewer are highly appreciated.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.

org/10.1016/j.soildyn.2020.106471.

References

[1] Robertson PK. Interpretation of cone penetration tests – a unified approach. Can
Geotech J 2009;46(11):1337–55.
[2] Rix GJ, Stokoe KH. Correlation of initial tangent modulus and cone penetration
resistance. In: Calibration chamber testing; 1991. p. 351–62.
[3] Mayne PW, Rix GJ. Correlations between shear wave velocity and cone tip
resistance in clays. Soils Found 1995;35(2):107–10.
[4] Hegazy YA, Mayne PW. Statistical correlations between vs and cone penetration
data for different soil types. In: Proceedings cone penetration testing (CPT’95), vol.
2. Swedish Geotechnical Society; 1995. p. 173–8.
[5] Mayne PW. In situ test calibrations for evaluating soil parameters. In: Proc.
Characterization and engineering properties of natural soils II; 2006. Singapore.
[6] Andrus RD, Mohanan NP, Piratheepan P, Ellis BS, Holzer TL. Predicting shear-wave
velocity from cone penetration resistance. In: Proc., 4th inter. Conf. On earthq.
Geotech. Eng; 2007. Thessaloniki, Greece.
[7] Wair BR, DeJong JT, Shantz T. “Guidelines for estimation of shear wave velocity
profiles”, pacific earthquake engineering research center. 2012.
[8] American Society of Civil Engineers. Minimum design loads for buildings and other
structures. 2006. Reston, vol. A.
[9] EN. Eurocode 8: design of structures for earthquake resistance – Part 1: general
rules, seismic actions and rules for buildings. 2004.
[10] McGann CR, Bradley BA, Taylor ML, Wotherspoon LM, Cubrinovski M.
Development of an empirical correlation for predicting shear wave velocity of
Christchurch soils from cone penetration test data. Soil Dynam Earthq Eng 2015;
75:66–75. https://doi.org/10.1016/j.soildyn.2015.03.023.
[11] Vasileiadis M. Region-specific CPT-vs empirical correlation. Arup technical note
229746_032.0_NOT136__Rev.0.01. 2015.
Fig. 18. VS30 for Groningen CPT database, obtained from the new empirical [12] Kulhawy FH, Mayne PH. “Manual on estimating soil properties for foundation
equation Eq. (4) to convert CPT parameters to VS and next convert to VS30. Each design”, Report EPRI-EL-6800. Palo Alto, California: Electric Power Research
dot represents a CPT of at least 29 m depth. The background shows the VS30 Institute; 1990.
[13] Kruiver PP, Kloosterman FH, de Lange G, Doornenbal P. Background document
zonation model [15] from the Groningen Ground Motion model version 6.
NAM database of subsurface information - version date of database - 29 March

14
P.P. Kruiver et al. Soil Dynamics and Earthquake Engineering 140 (2021) 106471

2018. 2018. https://nam-onderzoeksrapporten.data-app.nl/reports/download/ [19] NEN (NEderlandse Norm). National annex to NEN-EN 1997-1 Eurocode 7:
groningen/en/c39c93ac-9fdf-46ea-89a1-2e6c08e0ac34. geotechnical design - Part 1: general rules. Delft, the Netherlands: NEN; 2019.
[14] Noorlandt R, Kruiver PP, de Kleine MPE, Karaoulis M, de Lange G, Di Matteo A, [20] van der Linden TI, de Lange DA, Korff M. Cone penetration testing in thinly inter-
et al. Characterisation of ground motion recording stations in the Groningen gas layered soils. In: Proceedings of the institution of civil engineers- geotechnical
field. J Seismol 2018;22:605. https://doi.org/10.1007/s10950-017-9725-6. engineering, vol 171; 2018. p. 215–31. https://doi.org/10.1680/jgeen.17.00061.
[15] Kruiver PP, van Dedem E, Romijn R, de Lange G, Korff M, Stafleu J, et al. An 3.
integrated shear-wave velocity model for the Groningen gas field, The Netherlands. [21] de Lange DA. CPT in thinly layered soils – validation tests and analysis for multi
Bull Earthq Eng 2017;15(9):3555–80. https://doi.org/10.1007/s10518-017-0105- thin layer correction. 2018. https://nam-feitenencijfers.data-app.nl/download/ra
y. pport/3a1f937a-799a-4921-a64f-eac65acf14f7?open=true.
[16] Stafleu J, Dubelaar CW. Product specification Subsurface model GeoTOP. TNO report [22] R Core Team. R: a language and environment for statistical computing. Vienna,
2016-R10133 v1.3. TNO (utrecht, The Netherlands). 2016. www.dinoloket.nl/en Austria: R Foundation for Statistical Computing; 2018. https://www.R-project.org/
/want-know-more. .
[17] Douglas JB, Olsen RS. Soil classification using electric cone penetrometer. In: [23] Nikkels J. Geotechnisch veld werk betreffende seismische sonderingen Noordzijde
Symposium on cone penetration testing and experience. St. Louis: Geotechnical Eemskanaal. Fugro report 1316-0164-000 d. 2016.
Engineering Division, ASCE; 1981. p. 209–27. [24] Rodriguez-Marek A, Kruiver PP, Meijers P, Bommer JJ, Dost B, van Elk J, et al.
[18] Robertson PK, Cabal KL. Guide to cone penetration testing for geotechnical A regional site-response model for the Groningen gas field. Bull Seismol Soc Am
engineering.. 6th Signal Hill, CA: Gregg Drilling & Testing, Inc; 2015. 2017;107(5):2067–77. https://doi.org/10.1785/0120160123.

15

View publication stats