Evaluation of Charpy-V Transition Curve Fitting Based On Distribution-Free Statistical Assessment

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

International Journal of Pressure Vessels and Piping 206 (2023) 105079

Contents lists available at ScienceDirect

International Journal of Pressure Vessels and Piping


journal homepage: www.elsevier.com/locate/ijpvp

Evaluation of Charpy-V transition curve fitting based on distribution-free


statistical assessment
Kim Wallin a, *, Wolfram Baer b
a
KW-Solutions Ltd., Heinäkuja 4A, 02760, Espoo, Finland
b
Federal Institute for Materials Research and Testing (BAM), Unter den Eichen 87, D-12205, Berlin, Germany

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

Keywords: A non-parametric, distribution-free, statistical assessment of 5 large Charpy transition curve data sets is used for
Charpy impact test the optimum fitting equations for smaller data sets. The assessment makes use of a combination of rank prob­
Transition curve ability and binomial probability analysis of the data. The original non-parametric assessment method is
Fitting
improved by combining upper and lower bound binomial estimates, thus removing a bias that exists in the
Non-parametric assessment
Statistics
original method. The non-parametric assessment is not suitable as a standard method because it requires too
Rank probability many data points to give a reliable result. It is, however, ideal as a research tool to examine transition curve
Binomial probability shape and scatter. Based on the assessment several recommendations for transition curve fitting can be made.

1.1. The Tanh transition equation

The most common function usedvto describe the transition curve is


the so-called Tanh (hyperbolic tangent) equation having the form of Eq.
(1), which is standardized for example, in ASTM E2215 [13].
1. Introduction
( ( ))
US − LS T − Tt50
KV = ⋅ 1 + tanh + LS (1)
Charpy impact test properties of ferritic steels show a transitional 2 C
behaviour with temperature from a lower shelf representing fully brittle
fracture to an upper shelf representing fully ductile fracture. When The Tanh equation has four parameters, two of which (upper shelf
Charpy impact test data is available over the whole transition region, the absorbed energy, US, and lower shelf absorbed energy LS) can usually be
data can be fitted by a sigmoidal function. Over the years, a many defined separately. The upper shelf value can be separately estimated
different functions and fitting procedures have been proposed [1–12]. from the analysis of the impact energies corresponding to 100 % ductile
The problem with all proposed procedures is that they make a pre­ fracture appearance (SFA = 100 %). In some cases, upper shelf is defined
determined assumption regarding the sigmoidal function and/or the as SFA ≥95 %, but this is not true upper shelf. The LS absorbed energy
data behaviour. Thus, a mismatch between the fit curve and the real represents pure brittle fracture and corresponds to the asymptotic lower
material behaviour to be represented can occur. This work looks at the bound which is no longer affected by temperature. If there is no data for
data without making any assumptions regarding the function type or the the lower shelf absorbed energy, LS can be conservatively fixed e.g., to 2
data behaviour. It makes use of a distribution-free statistical assessment J. The LS value is a weak function of yield stress. For normal strength
method. First an introductory of the most used Tanh transition equation steels 2 J is a realistic value, but for extra high strength steels the LS
is given describing two different fitting algorithms. Then the value increases [15]. The physical reason is that fracture requires some
distribution-free statistical assessment method is presented. It makes use plasticity at the notch root and such plasticity depends on the yield
of a combination of rank and binomial probability analysis. In the sec­ strength. Fixing LS and US leaves only two parameters to be fit, Tt50
ond part, the distribution free statistical assessment method is applied to which represents the temperature in the middle of the transition region
several large data sets and the results are compared to different and C which represents half of the width of the transition region. As
sigmoidal fits. Finally, based on large data set results, recommendations such, all four parameters have a physical meaning.
for the optimum fitting methods for small data sets are proposed. An example of a transition curve for a homogeneous forging is

* Corresponding author.
E-mail addresses: kim.wallin@kwsolutions.fi (K. Wallin), wolfram.baer@bam.de (W. Baer).

https://doi.org/10.1016/j.ijpvp.2023.105079
Received 11 September 2023; Received in revised form 9 October 2023; Accepted 16 October 2023
Available online 18 October 2023
0308-0161/© 2023 Elsevier Ltd. All rights reserved.
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Nomenclature SFA Shear fracture appearance


T Temperature in ◦ C
ai Fitting parameters in Böhme’s equation Tanh Hyperbolic tangent equation
atan Inverse tangent function Tt27J Temperature corresponding to 27 J
B Specimen thickness Tt40J Temperature corresponding to 40 J
C Parameter describing half width of the transition range in TtxJ Temperature corresponding to x J
terms of temperature T0 Scaling temperature in Burr equation
CVN Charpy-V notch Tt50 Temperature corresponding to energy (US + LS)/2 (middle
h Curvature parameter in Kohout’s equation of the transition range on the scales of both temperature
KV Absorbed energy for a Charpy V-notch test piece and absorbed energy, i.e. at the inflection point of a
LE Lateral expansion symmetric function)
LS Lower shelf absorbed energy Tmin
US Lowest temperature with US results
k Scaling parameter in Burr’s equation US Upper shelf absorbed energy
m Skewness parameter in Burr’s equation US Average upper shelf absorbed energy
N Sample size σ Standard deviation
n Number of data in analysis σKV Standard deviation of absorbed energy
neff Effective number of specimens below upper shelf and σUS Upper shelf absorbed energy standard deviation
above lower shelf σTt Standard deviation of temperature corresponding to Tt
nUS Number of specimens corresponding to upper shelf σTtxJ Standard deviation of temperature corresponding to x J
P Probability σY Yield strength
PB0.5 Median Binomial probability
PR0.5 Median Rank probability

presented in Fig. 1 [14]. The scatter may at a first glance appear large, then the temperature only specifies where on the curve the specimen
but this is attributed to the large number of tests (200). Normally CVN lies. In this case the physical argument changes to one where parallel
transition curves contain around 30 specimens but may be as few as only transition curves differ by some temperature amount. In this case a
10–12 specimens, which is common for example, in nuclear surveillance temperature-based fit is more appropriate. This latter assumption is
programs [13]. When the Tanh equation is fitted to such a small data set, supported by the fact that data scatter is more consistent with respect to
the scatter appears less than what it really is. temperature than energy as shown in Fig. 2. The scatter with respect to
There are several different techniques by which one can fit the Tanh energy was determined at each test temperature with 20 specimens per
equation to CVN impact data. They range from quite sophisticated to temperature. The scatter with respect to temperature was obtained by
simple least square methods. Most often, the Tanh equation is fitted by a ordering the 200 data points by energy and dividing the data in 10
least squares fit with relation to absorbed energy KV, lateral expansion groups with 20 specimens each. The scatter and mean temperature were
LE or shear fracture appearance SFA. This is the method described in then determined for each group.
different standard procedures like ASTM E2215 [13] and NIST [11], but The scatter in terms of KV (LE, SFA) is not constant, but varies over
the method is not always optimal for CVN fitting. Another possible the transition region, being largest in the centre of the transition region
parameter, on which the least square analysis can be based, is the and smallest in the lower and upper shelves. The scatter in terms of
temperature. Normally it is assumed that temperature is defined without temperature, however, is nearly constant, as seen from Fig. 2, thus fa­
scatter and all the variability occurs in terms of energy. This is the voring parallel transition curves. The least square fitting methods as­
common argument for using the energy-based fitting. However, if we sume a constant scatter over the whole fitting range. This means that too
assume that each specimen is connected to a specific transition curve, much weight is put on the data in the middle of the transition, when
fitting in terms of energy. One can try to estimate the scatter along the

Fig. 1. Typical CVN transition curve for nuclear grade forging based on 200 Fig. 2. Standard deviation of data presented in Fig. 1, showing the difference in
tests [14]. The temperature and energy based fits are explained below. scatter behaviour in terms of energy and temperature [14,15].

2
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

transition curve and use this to weigh the individual energy values, but The role of the dummy parameter is to ensure that the Tanh equation
this is practically impossible for a small data set. reaches upper shelf in the correct temperature region. This is especially
Fitting in terms of temperature removes this problem. It automati­ important for steep transition curves. For large data sets, the dummy
cally ensures the correct weighing of each data point regardless of test parameter has only limited effect, but for small data sets the use of the
temperature. The only problem is that very close to LS and US, the dummy parameter is important.
scatter in terms of temperature increases. Therefore, these regions The steepness parameter C is obtained from Eq. (4) [11] and the
should not be included in the fit, except for estimating LS and US. The normalization temperature Tt50 from Eq. (5) [2].
inverted Tanh equation has the form shown in Eq. (2). ( ) ( )
∑n
US− LS
∑ n ∑n
US− LS
( ) n⋅ Ti ⋅ln KV − LS
− 1 − T i ⋅ ln KV − LS
− 1
C KV − LS i i
T = Tt50 + ⋅ln (2) C = 2⋅ (i=1 ( ))2
i=1 i=1
( )2 (4)
2 US − KV ∑n
US− CV min ∑ n
US− LS
ln KV i − CV min
− 1 − n⋅ ln KVi − LS
− 1
Contrary to the energy-based fit, which requires a numerical fitting i=1 i=1

procedure, the temperature-based fitting algorithm can be expressed as ( )



n ∑
n
a closed form solution. The principle of the fitting algorithm is presented Ti + C2⋅ ln KV
US− LS
− 1
i − LS
in Fig. 3. The figure includes estimates of the 5 % and 95 % scatter bands Tt50 = i=1 i=1
(5)
as described further on. n
The fitting is divided into two parts. First, the upper shelf absorbed
in Equations (3)–(5), n is the number of data points with 0.1• US ≤ KVi
energy, and its standard deviation are determined with the measured
≤ 0.85• US plus the dummy datum. The standard deviation is then
values of all test pieces showing 100 % ductile fracture surface using Eq.
obtained from Eq. (6) where the dummy data is omitted. Depending on
(3) [14].in Eq. (3), nUS is the number of results corresponding to upper
the data, the region for data selection may vary. A suitable range is
shelf and USi is the upper shelf absorbed energy for each individual
result. The lower shelf absorbed energy LS is normally fixed at some 1.65•σUS ≤ KVi ≤ US-1.65•σUS.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
realistic value based on test data, or if the data allows an estimate of the √ i=1 ( ( ))2
√∑
asymptotic lower bound, this should be used. Simply using the average σTt = √ Ti − Tt50 − ⋅ln
C US − LS
− 1 (6)
of all specimens with SFA = 0 % may provide too high estimates of the n
2 KVi − LS
asymptotic LS, because LS corresponds mathematically to a limiting
The energy and temperature-based Tanh fits are shown in Fig. 1. For
lower bound. Two parameters (Tt50 and C) still need to be estimated. The
this large data set, the least square fits are not sensitive to the use of
fitting makes use of the data points corresponding to the region where
energy or temperature, but the temperature-based fit can also provide an
the temperature variation is almost uniform. This region is defined as
estimate of the scatter. Fig. 1 shows the 5 % and 95 % scatter bands
approximately 0.1• US ≤ KVi ≤ 0.85• US. The region is not symmetrical,
estimated based on a combination of Eqs. (3) and (6), but the scatter
because the upper shelf includes scatter and the lower shelf does not. If
bands can also be established for other fractiles, like ±1σ.
there are insufficient data in this region, the limits may be changed to
Normally, the number of specimens represented by a transition curve
include more data. In addition to the data belonging to this region, one
is much smaller than 200. Also, transition curves are usually used to
additional dummy datum is added. The dummy datum is fixed as having
estimate a specific transition temperature like Tt27J or Tt40J. In this case
the absorbed energy value KV = 0.97• US and corresponding to the
the selection of fitting parameters may become more important.
lowest temperature where upper shelf data has been obtained (Tmin US ). To examine the validity of the two fitting parameters, the data set in
Fig. 1 was analyzed with a non-parametric distribution-free statistical
procedure.

1.2. Non-parametric analysis

The distribution-free statistical assessment method is based on a


combination of Rank and Bimodal probability estimates. The method
can be used to determine which kind of distribution is most appropriate
for the data or it can be used directly to develop statistically defined
lower and upper bound estimates without requiring knowledge of the
actual distribution. The principle of the distribution-free statistical
assessment method is shown in Fig. 4. In the Rank probability method,
the data at any one temperature are ordered by rank. The probability of
the i:th value is then equal to or less than xi. The Binomial probability
method, shown in Fig. 4, estimates the probability that, for a data set
ordered by rank, above or below a certain temperature, the lowest i
values are equal to or lower/higher than xi. The definitions may appear
similar, but they do contain two significant differences. First, the Rank
probability is connected to a specific test result, whereas the Binomial
Fig. 3. Schematic presentation of modified temperature-based fitting algo­ probability is connected to some freely chosen criterion xi. Second, the
rithm [14].
Rank probability exists for the values i = 1 … n, whereas the Binomial

probability exists for the values i = 0 … n. These differences make the
nUS
USi
US = i=1
methods suitable for producing two independent descriptions of the data
nUS
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (3) scatter. The method has been described before in Refs. [16,17] where
√∑
√nUS fatigue and fracture toughness data have been analyzed. Previously [16,
√ US2i
√ 17], the Binomial probability method has only been applied to data
σUS = i=1 − US2
nUS either below or above a certain criterion, but this leads to a slightly
biased estimate. This is because the test temperatures are not random
but limited to specific values. This causes temperature intervals where

3
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

estimates are obtained by taking the average of the upper and lower
bound estimates. Note that the probabilities in the upper and lower
bound estimates are reversed with respect to probability.
Fig. 7 shows as an example the estimates of the ±1σ fractiles. The
upper and lower bound Binomial estimates show a clear bias. The points
marked “combined” are the averages of the Binomial estimates and the
estimates based on the Rank probability. As seen, they form together an
estimate with very little scatter. The combined estimates were fitted by a
so-called Burr equation [1,6,7,10] which is quite flexible. For this large
data set and material, the Burr equation fits the non-parametric esti­
mates satisfactorily.
The Burr equation used here has the form given by Eq. (9).
(US − LS)
KV = [ ( )]m + LS (9)
1 + exp − T− kT0

If fixed, US and LS can represent the upper shelf and lower shelf
absorbed energies, but for a better overall fit to the combined estimates,
Fig. 4. Principle of rank and binomial analysis. they were used as fitting parameters. Thus, they do not represent true
upper or lower shelf energies in this case. Unfortunately, the parameters
the Binomial probability is constant. The size of the bias is then T0, k and m lack any physical meaning. From this respect the Tanh
dependent on the size of the intervals between test temperatures. By equation is physically sounder.
combining both values above a certain energy up to a selected temper­ Fig. 8 shows a comparison of the non-parametric estimates for me­
ature and values below a certain energy down to a selected temperature dian and ±1σ fractiles and estimates based on the Tanh equation using
and taking their average, the bias is avoided. both temperature and energy as fitting parameter. The median estimates
One of the most common simple analytical median Rank probability do not differ significantly, but the energy-based median fit seems to
estimates has the form given in Eq. (7) [15,16]. overestimate the lower transition region with respect to the non-
parametric analysis. The temperature-based Tanh equation ±1σ frac­
PR0.5 ≈
i − 0.3
(7) tile estimates show good agreement with the non-parametric estimates.
n + 0.4 It should be noted that the data set is sufficiently large to provide a high
For the Binomial median probability estimate a similar equation has accuracy for the non-parametric analysis. It was established that, in
been developed, Eq (8) [15]. order to have a good reliability in the estimates, the non-parametric
analysis requires the average number of specimens both for the Rank
i + 0.684
PB0.5 ≈ (8) and Binomial probability to be n and Σn ≥ 8 [17]. This is clearly fulfilled
n + 1.368
with this large data set. The n and Σn varied in all cases between 20 …
in Eq. (8), i corresponds either to the upper bound or lower bound 50 making the result quite reliable.
number of data and n is the corresponding total number for upper or Below the analysis is applied on four additional comparatively large
lower bound analysis. CVN data sets originally published in Ref. [12].
The resulting Rank probability diagram is shown in Fig. 5. The figure
includes the 5 %, -1σ, median, +1σ, and 95 % lines. The corresponding 2. Non-parametric analysis of different strength steels
Binomial probability diagrams are shown in Fig. 6. These figures also
include the 5 %, -1σ, median, +1σ, and 95 % lines. The combined Four different structural steels with comparatively large CVN data
sets were selected for the analysis. The data is taken from Ref. [12]. The
steels corresponded to the EN 10025 [18] standard designations:
S235JR, S355 N, S460 N, and S960QL. The yield strengths varied from
282 MPa for the S235JR steel to 1015 MPa for the S960QL steel. Charpy
tests were carried out according to ISO 148-1 [19] and ISO 14556 [20]
using an instrumented ISO type striker with a 2 mm radius of the striking
edge. The capacity of the machine was 300 J. The impact test results for
all four steels are shown in Fig. 9. Besides performing a non-parametric
analysis of the data, also more conventional data analysis (Tanh fit) was
performed. In Ref. [12], the upper shelf was estimated as the average
energy of all specimens at temperatures where all specimens showed
100 % ductile fracture. Here, all specimens showing 100 % ductile
fracture, regardless of temperature were used to estimate the upper shelf
energy US. Also, in Ref. [12], the lower shelf was taken as the average of
all specimens showing 0 % ductile fracture. Here a more conservative
value, based on a visual estimate of the apparent asymptotic value
corresponding to temperature insensitivity was used for the lower shelf
energy LS.
The Rank probability analyses for the four steels are shown in Fig. 10.
There should be preferably a minimum of 8 test results per temperature
to obtain a reliable analysis. Unfortunately, only the S235JR steel to a
significant degree fulfils this requirement. The other steels have less tests
per temperature and the Rank probability estimate is not quite accurate.
Fig. 5. Rank probability analysis of the data in Fig. 4.

4
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 6. Binomial probability analysis of the data in Fig. 4.

Fig. 7. Distribution-free ±1σ estimates for the data in Fig. 4, showing bias in one-sided Binomial probabilities and the beneficial effect of combining upper and lower
bound values. The combined estimates are the averages of the Binomial estimates combined with the Rank estimates.

It is evident that it is better to test more specimens at fewer temperatures


if a non-parametric analysis is intended.
The Binomial probability analyses for the four steels are shown in
Figs. 11–14. The probability trends in the figures do not in all cases show
smooth trends like in Fig. 6. Instead, there are several clear disconti­
nuities in the trends, indicating too few data points to obtain a reliable
probability estimate. This will have an effect on the overall non-
parametric estimate reliability.
The resulting median and ±1σ estimates are presented in Fig. 15.
The variation in the ±1σ estimates is larger than for the 15Ch2MFA
steel. This is because the 15Ch2MFA steel has nearly three times more
specimens, which makes the non-parametric estimate more accurate.
Due to the larger variability in the non-parametric estimates, their
curve fits are not as good as for the 15Ch2MFA steel. One thing is clear
from the analysis. The temperature dependence does not follow a sym­
metric function. Thus, an equation like the Tanh equation is not ideal.
The Burr equation (8) is better suited to describe the data.
Taking half of the difference between the ±1σ estimates, provides an
estimate of the standard deviation as a function of impact energy. En­
Fig. 8. Comparison between distribution-free analysis of data in Figs. 1 and 4 ergy is used instead of temperature to obtain overlapping graphs. Fig. 16
and Tanh fits based on energy and temperature. The non-parametric fit is shows the resulting non-parametric standard deviation estimates as a
described with the Burr equation fit to the combined estimates. function of impact energy in relation to the individual US values. En­
ergies below [LS-0.1•US] and above 0.9•US are affected by the tem­
perature independent LS and US regions and should be omitted, since

5
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 9. CVN data for the four different steels [12].

the scatter in these regions is connected to excess uncertainty as seen in 2. Discussion


Fig. 16. The overall standard deviation is not strongly affected by energy
level and the value is close to 10 ◦ C ± 5 ◦ C, similar to the 15Ch2MFA The whole transition curve is rarely used in engineering applications.
steel. Considering that the uncertainty in the non-parametric ±1σ esti­ Instead the curve is used to estimate a temperature corresponding to
mates for these data sets is of the order of ±10 ◦ C, as seen from the some selected energy level. Typical energy levels used in materials
scatter in Fig. 16, the assumption of constant scatter with respect to specifications are e.g., 27/28 J and 40/41 J, but other values are also
temperature is supported by the data. used for varioius applications. In some applications the energy level is
Similar, to the Tanh equation, the Burr equation can also be inverted connected to the materials yield strength, like 0.1•σY [21].
so that the fitting is done in terms of temperature. The equation can be The 27 J and 40 J temperatures estimated by the energy- and
expressed as Eq. (10). temperature-based Burr fits for the four structural steels were compared
{( )1/m } in Fig. 18 with the corresponding estimates based on the median non-
T = T0 − k⋅ln
US − LS
− 1 (10) parametric estimates. The variation of the estimate variation is of the
KV − LS order of ±5 ◦ C, but the temperature-based estimate tends to be slightly
conservative, whereas the energy-based estimate tends to be slightly
Estimating US and LS separately from the experimental data, three
unconservative. Fig. 18 also includes the comparison for the 15Ch2MFA
fitting parameters are left, T0, k and m. Similar, to the Tanh equation, it
steel (triangles). Here the variation is smaller, but the trend remains the
is advisable to leave out data points too close to US and LS, because the
same.
variation in terms of temperature becomes too large in these regions.
If test temperatures are properly spaced, medium-sized data sets
The actual fitting can be performed with any non-linear curve fitting
containing 30 to 50 results will be suitable to establish a reasonable
program.
estimate of a whole transition curve. The Burr equation can be applied to
Fig. 17 shows the comparison of energy- and temperature-based Burr
obtain satisfactory fits of data sets of such size. Alternatively, Eq. (11)
equation fits to the actual test results and the non-parametric median
provides an example of another fitting option. Eq. (11), which was
Burr fit. There are no major differences between the three fits. In some
proposed by Kohout [10] and applied to the data sets in Ref. [12], is
energy regions the energy-based fit follows better the non-parametric
basically a symmetric function like Tanh. But Eq. (11), with one more
estimate, and in other energy regions the temperature-based fit does
fitting parameter than Tanh, is more flexible regarding fitting capacity
the same. As a general trend, the temperature-based fit is more conser­
than the simple Tanh equation. It was proposed in Ref. [12] that Eq. (11)
vative than the energy-based fit.

6
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 10. Rank probability analysis of the four steels.

Fig. 11. Binomial probability analysis of steel S235JR.

7
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 12. Binomial probability analysis of steel S355 N.

Fig. 13. Binomial probability analysis of steel S460 N.

Fig. 14. Binomial probability analysis of steel S960QL.

8
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 15. Non-parametric median and ±1σ estimates for the four structural steels. The non-parametric fit is described with the Burr equation fit to the median and
±1σ estimates.

may provide better fits than the Tanh equation.


( ( ) )1/h
US + LS US − LS 2 π |T − Tt50 | h
KV = + ⋅ SGN(T − Tt50 )⋅ ⋅atan ⋅
2 2 π 2 C
(11)
Eq. (11) requires the data sets to be close to symmetrical to be reli­
able. The result of using the Kohout equation on the investigated ma­
terials is shown in Fig. 19, with a similar form as for the Burr equation in
Fig. 18. Since it is mathematically difficult to invert Eq. (11), only the
energy-based fit is shown in Fig. 19. For the symmetric 15Ch2MFA steel
data (triangle), the Kohout equation provides a similar accuracy as the
Burr equation, but for the four more non-symmetric data sets, the ac­
curacy of the Kohout equation is lower than that of the Burr equation.
Other more complicated equations have been suggested [2,8], but
their value with respect to more reliable fits is questionable. With
additional fitting parameters, the fit to the individual data points be­
comes better, but this may cause “overfitting”, which may lead to wrong
conclusions simply due to the uneven scatter of the test results. The
smaller the data set the larger is the danger of overfitting. With small
data sets one should always try to use an equation with a minimum
Fig. 16. Non-parametric standard deviation estimates for the four structural
number of fitting parameters. As an example, we can consider an
steels. The region best describing the overall scatter is between [LS+0.1•US]
extremely flexible equation proposed by Böhme [8]. It contains 6 fitting
and 0.9•US (vertical lines).
parameters and has the form shaown in Eq. (12).
( )
a1 ⋅(T − a5 ) a2 ⋅(T − a5 )2 T − a5
KV = a0 + + + a 3 ⋅tanh (12)
a4 a24 a4

9
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 17. Comparison of energy- and temperature-based Burr equation fits to the actual test results and the non-parametric median Burr fit for the four steels.

Fig. 19. Comparison between transition temperature estimations based on the


Fig. 18. Comparison between transition temperatures estimated by the energy-
Kohout equation and the corresponding estimates based on the median non-
and temperature-based Burr fits and the corresponding estimates based on the
parametric Burr fit for the four structural steels. The triangles refer to the
median non-parametric Burr fit for the four structural steels. The triangles refer
large data set comprising 200 specimens of the reactor pressure vessel steel.
to the large data set comprising 200 specimens of the reactor pressure
vessel steel.

10
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 20. Two random subsets from the 15Ch2MFA steel data set fitted with Eq. (12).

The fitting parameters a0 … a5 do not have physical meaning. Thus, However, for small data sets with less than 30 test results, the benefit
they cannot be connected to LS or US. As an example of the flexibility of of using a more flexible equation is questionable, and only a rough es­
Eq. (12), two subsets from the 15Ch2MFA steel data set were randomly timate of the transition curve can be obtained. For such data sets the
selected by taking one random value at each temperature, thus obtaining simple Tanh equation may be a suitable choice.
two subsets with 10 data points each. The data was fitted with Eq. (12). As to whether the equation should be fit with respect to energy or
As seen in Fig. 20, the equation is extremely flexible and provides a very temperature, a previous study [15] compared 28 J and 41 J tempera­
good fit to the data. The problem is that the data sets are too small to tures fitted with both methods for small data sets with 10 data points
give a reliable description of the transition curve. Even though the data each. The data was taken by randomly sampling data from the
subsets represent the same material, the resulting fits would not indicate 15Ch2MFA steel data set, similarly to the data shown in Fig. 20. The
this. This is what is meant by “overfitting” in the case of small data sets. distributions of 28 J and 41 J temperatures estimated by the two
Based on an earlier assessment of the 15Ch2MFA steel [15], an methods are shown in Fig. 21. For both energy levels, the estimates
approximate equation for the standard deviation of different tempera­ based on temperature produced normally distributed values. The esti­
ture estimates corresponding to a given energy x has the form of Eq. mates based on energy showed tails that did not follow a normal dis­
(13). tribution. Due to this, the energy-based fits showed a larger scatter than
σT the temperature-based fits. The temperature-based fits standard devia­
σTt xJ ≈ ( )0.3 (13) tion is well in line with Eq. (12). For the larger data sets, the
neff − 4
energy-based fits tend to produce lower temperatures than the
The parameter neff relates to the number of data points below US and temperature-based fits. The best estimate non-parametric values are
above LS. For the four steels, neff varies in the range 38–57. Considering Tt28J = − 60 ◦ C and Tt41J = − 53 ◦ C. The respective median values for
that σTt for all four steels is close to 10 ◦ C, the estimate for σTtxJ is temperature-based fits are − 59 ◦ C and − 51 ◦ C, whereas the corre­
approximately 3 ◦ C. For the 15Ch2MFA steel where neff is 125, the es­ sponding values for the energy-based fits are − 62 ◦ C and − 52 ◦ C. Even
timate for σTtxJ is approximately 2 ◦ C. This is fully in line with the trend though the differences are small, the trend is the same. The
seen in Fig. 18. temperature-based fits are slightly conservative, whereas the
Based on the above results, it is recommended to use the Burr energy-based fits are slightly non-conservative.
equation fitted with respect to temperature as given by Eq. (9) for larger In order to examine more closely the effect of overfitting on 28 J and
data sets with approximately 30 test results or more. 41 J temperature estimates, Eq. (12) was used to analyze the same data

Fig. 21. Distribution of temperatures (Tt28J and Tt41J) estimated from random sub-sets with 10 data points from the large 15Ch2MFA data set with 200 data,
analyzed with energy- and temperature-based Tanh equations [15].

11
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

Fig. 22. Distribution of temperatures (Tt28J and Tt41J) estimated from random sub-sets with 10 data points from the large 15Ch2MFA data set with 200 data,
analyzed with energy-based Böhme [8] and temperature-based Tanh equations [15]. Also included are the corresponding distributions for the inverted Burr equation.

sets as the ones leading to Fig. 21. Fig. 22 shows the comparison between 4) Small data sets should not be fitted with equations containing too
Eq. (12) estimates and the temperature-based Tanh estimates. The effect many parameters, since it will lead to overfitting and more scatter in
of overfitting is seen as an increased uncertainty in the estimate. The specific temperature estimates, like Tt27J and Tt40J.
result is even worse than for the energy-based Tanh estimates in Fig. 21.
The same data sets were also analyzed with the inverted Burr equation Author statement
and the resulting distributions are also included in Fig. 22. The inverted
Burr equation provides almost the same distributions as the The corresponding author is responsible for ensuring that the de­
temperature-based Tanh estimates. This indicates that the inverted Burr scriptions are accurate.
equation can also be used for relatively small data sets without danger of
overfitting.
A practical additioinal advantage of the temperature-based fit Declaration of competing interest
compared to the energy-based fit is that the obtained formula easily
allows calculating transition temperatures at desired KV values. The The authors declare that they have no known competing financial
more complicated search of target values using solver routines - as interests or personal relationships that could have appeared to influence
needed with energy-based fit formulae is unnecessary. the work reported in this paper.

3. Summary and conclusions Data availability

A non-parametric, distribution-free, statistical assessment of 5 large The authors do not have permission to share data.
CVN transition curve data sets has been applied to look for optimum
fitting equations for smaller data sets. The assessment method makes use Acknowledgements
of a combination of rank probability and binomial probability analysis
of the data. The non-parametric assessment requires comparatively large The outstanding experimental work of the Fracture Mechanics and
data sets with at least 100 data points to obtain reliable descriptions of Structural Integrity Laboratory at BAM Federal Institute for Materials
both curve shape and scatter. With data sets consisting of 50–100 data Research and Testing, Berlin, Germany, which performed the compre­
points, the non-parametric assessment can be used to estimate the me­ hensive test program to establish the material data base for the struc­
dian behaviour of the data, but the estimate of the scatter contains un­ tural steels discussed in this paper and in Ref. [12] is gratefully
certainty. The original non-parametric assessment method was acknowledged.
improved by combining upper and lower bound binomial estimates,
thus removing a bias that existed in the original method. The non- References
parametric assessment is not suitable as a standard method because it
requires too many data points to provide a reliable result. It is, however, [1] I.W. Burr, Cumulative frequency functions, Ann. Math. Stat. 13 (2) (1942)
215–232.
ideal as a research tool to examine transition curve shape and scatter.
[2] G. Hofer, C.C. Hung, Ü. Günes, A mathematical function for the description of
Based on the assessment the following recommendations can be made: Charpy impact tests, Z. für Werkst. 8 (1977) 109–111.
[3] W. Oldfield, Fitting curves to toughness data, Journal of Testing and Evaluation,
1) It is recommended that the CVN transition curve fitting is perfomed JTEVA 7 (6) (1979) 326–333.
[4] F.W. Stallman, Evaluation and uncertainty estimates of Charpy—impact data, in:
based on temperature since the scatter in terms of temperature is Proceedings of the Fourth ASTM-EURATOM Symposium on Reactor Dosimetry,
more constant than for energy (or LE or SFA). National Bureau of Standards, Gaithersburg, Maryland, 1982. March 22-26.
2) For data sets with more than 30 results, fitting may be based on the [5] P. Urwank, Unambiguous curve fitting and error estimation for Charpy impact test
data of reactor pressure vessel steels, suitable for small number of samples, J. Nucl.
Burr equation fitted in terms of temperature. Mater. 161 (1989) 24–29.
3) For data sets with less than 30 results, fitting can simply be based on [6] J.L. Helm, The interpretation of Charpy impact test data using hyper-logistic fitting
the standard Tanh equation fitted in terms of temperature, even functions, in: 17:th Symposium on Effects of Radiation on Materials, ASTM special
technical publication, 1996, 1270.
though the Burr equation fitted in terms of temperature can also be [7] P.L. Windle, M. Crowder, R. Moskovic, A statistical model for the analysis and
used with sufficient confidence, if there are more than 10 results. prediction of the effect of neutron irradiation on Charpy impact energy curves,
Nucl. Eng. Des. 165 (1–2) (1996) 43–56.
[8] W. Böhme, W. Schmitt, Comparison of results of instrumented Charpy- and min-
Charpy tests with different RPV-steels, in: Small Specimens Test Techniques, ASTM

12
K. Wallin and W. Baer International Journal of Pressure Vessels and Piping 206 (2023) 105079

STP 1329, Corwin WR, Rosinski ST, Van Walle E., American Society for testing and quality system tool applicable for steel producers and steel users of heavy plates,
materials, 1998. profiles and weldments. Technical steel research, European commission, research
[9] L.W. Cao, S.J. Wu, P.E.J. Flewitt, Comparison of ductile-to-brittle transition curve fund for coal and steel, EUR 22454 (2007).
fitting approaches, Int. J. Pres. Ves. Pip. 93–94 (2012) 12–16. [15] K. Wallin, Fracture Toughness of Engineering Materials – Estimation and
[10] J. Kohout, Various Regression Functions Fitting Transition Curves, Internal Report Application, EMAS Publishing, Warrington, 2011.
for the Ministry of Defense of the Czech Republic, 2012. [16] K. Wallin, Distribution free statistical assessment of scatter and size effects in the
[11] E. Lucon, J. Splett, A. Koepke, D. Newton, NIST software package for obtaining Euro fracture toughness data set, Eng. Fract. Mech. 103 (1) (2013) 69–78.
Charpy transition curves, NIST Technical Note 2158 (2021), https://doi.org/ [17] K. Wallin, Simple distribution free statistical assessment of structural integrity
10.6028/NIST.TN.2158. material property data, Eng. Fract. Mech. 78 (9) (2011) 2070–2081.
[12] W. Baer, P. Wossidlo, B. Abbasi, On the question of how to analyze and apply [18] prEN 10025-1, Hot Rolled Products of Structural Steels – Part 1, General Technical
Charpy pendulum impact test results correctly - shortcomings in corresponding Delivery Conditions, 2011.
procedures and standards, JTE 50 (2022), https://doi.org/10.1520/JTE20210781. [19] ISO 148-1, Metallic Materials – Charpy Pendulum Impact Test – Part, vol. 1, Test
[13] ASTM E2215 Standard Practice for Evaluation of Surveillance Capsules from Light- Method, 2016.
Water Moderated Nuclear Power Reactor Vessels. [20] ISO 14556, Metallic Materials – Charpy Pendulum Impact Test – Instrumented Test
[14] P. Langenberg, J. Buchholz, A. Völling, W. Bleck, P. Balladon, K. Wallin, Method, 2015.
P. Nevasmaa, A. Laukkanen, QUALYTOUGH - development of a fracture toughness [21] NORSOK, STANDARD N-004: Design of Steel Structures, 2013.

13

You might also like