Chan, A. H. C. (2006). Géotechnique 56, No.

1, 73–76

DISCUSSION

Determination of the coefficient of consolidation using a least squares

method
A . H . C . C H A N ( 2 0 0 3 ) . G é o t e c h n i q u e 5 3 , N o . 7 , 6 7 3 – 6 7 8

I. G. Doran and J. D. McKinley, Queen’s University Belfast Fig. 4, which shows the difference between the approximate
The paper proposes the use of an approximation, sug- value of U and the exact value, relative to the exact value, as a
gested by Hansen, to the exact analytical solution of the function of time factor, for several different approximations.
Terzaghi consolidation equation, as the basis for the determi- This includes the series solution (equations (4) and (5) in the
nation of the coefficient of consolidation from the results of paper) truncated at different values of the index m. In this
a consolidation test using a least squares error adjustment discussion, ‘exact’ values have been calculated from the series
procedure. This is an area of considerable practical interest. solution summing up to m ¼ 10 000. This is straightforward if
The author seems to be unaware of the paper by Fox (1948), the series solution is embedded in a user-defined function in
in which it is shown that the solution of the Terzaghi Microsoft Excel using a FOR . . . NEXT loop.
consolidation equation is given to a high degree of accuracy The estimation of the consolidation parameters using Fox’s
by the approximation approximation can be achieved on a Microsoft Excel spread-
rffiffiffiffiffiffiffiffi sheet using the Solver tool to determine the curve fit that
4Tv gives the least square error (Doran & McKinley, in prepara-
U¼ 0 < Tv < 0:2 (19a)
ð tion), using a simple IF(. . .) function. We have also applied
8 ð2 Solver to the series solution to the Terzaghi consolidation
U ¼ 1
2 e
4 Tv Tv . 0:2 (19b) equation, as given by equations (4) and (5) in the paper, and
ð
found that in practice about five terms give sufficient
In fact, equations (1) and (6) of the paper are the same as accuracy for the interpretation of typical consolidation test
equations (19a) and (19b) above, respectively, although Fox results. However, in our experience manipulating even this
proposed that the transition point was Tv ¼ 0.2, which restricted number of terms is a tedious process unless the
corresponds to U  50% and not the U  60% (i.e. Tv  series solution is embedded in a user-defined function in
0.286) stated in the paper. Corresponding values of U and T Microsoft Excel, whereas the use of Fox’s approximation is
are shown in Table 6, which is an extension of Table 1 in straightforward and gives satisfactory results.
the paper. Clearly, the values calculated from Fox’s approx- Figure 4 suggests that improved versions of Fox’s approx-
imation are sufficiently accurate for any likely application, imation can readily be established by the incorporation of
and are also significantly more accurate than the Hansen extra terms from the series solution. Taking just the first
approximation used in the paper. term of the series, m ¼ 0, a more accurate representation of
The accuracy of Fox’s approximation is further illustrated in Fox’s approximation is

Table 6. Comparison of various approximate equations with the series solution to the
one-dimensional Terzaghi consolidation equation

Series solution, m ¼ 10 000 Hansen solution Fox solution

Tv U U Diff. U Diff.

0.001 0.0357 0.0355
0.0002 0.0357 0.0000

0.010 0.1128 0.1122
0.0006 0.1128 0.0000
0.020 0.1596 0.1587
0.0008 0.1596 0.0000
0.030 0.1954 0.1944
0.0010 0.1954 0.0000
0.040 0.2257 0.2245
0.0012 0.2257 0.0000
0.050 0.2523 0.2510
0.0013 0.2523 0.0000
0.060 0.2764 0.2749
0.0015 0.2764 0.0000
0.070 0.2985 0.2969
0.0016 0.2985 0.0000
0.080 0.3192 0.3174
0.0017 0.3192 0.0000
0.090 0.3385 0.3367
0.0019 0.3385 0.0000
0.100 0.3568 0.3548
0.0020 0.3568 0.0000
0.200 0.5041 0.5007
0.0034 0.5046 0.0005
0.300 0.6132 0.6094
0.0038 0.6134 0.0001
0.400 0.6979 0.6958
0.0021 0.6979 0.0000
0.500 0.7640 0.7647 0.0008 0.7640 0.0000
0.600 0.8156 0.8189 0.0034 0.8156 0.0000
0.700 0.8559 0.8608 0.0049 0.8559 0.0000
0.800 0.8874 0.8927 0.0052 0.8874 0.0000
0.900 0.9120 0.9166 0.0046 0.9120 0.0000
1.000 0.9313 0.9347 0.0034 0.9313 0.0000

73
74 DISCUSSION
Time factor, T

0·001 0·01 0·1 1 10

102

10

1021

1022

1023

1024

1025

1026

1027

1028

1029

10210

Terzaghi m 5 0 Terzaghi m 5 1 Terzaghi m 5 2 Terzaghi m 5 5

Terzaghi m 5 10 Terzaghi m 5 20 Fox Hansen

Fig. 4. Comparison of relative difference between various approximate equations with the series solution to the one-dimensional
Terzaghi consolidation equation

rffiffiffiffiffiffiffiffi
4Tv in practice lead to degenerate adjustments, and either ap-
U¼ 0 < Tv < 0:213 (20a)
ð proximation may be used with confidence.
8 ð2 The major difficulty that remains in the solution of the
U ¼ 1
2 e
4 Tv Tv . 0:213 (20b) general problem is the determination of the range of read-
ð
ings within a given set of test results that actually comply
and at the transition point U  52.1%. For this approxima- with Terzaghi consolidation theory, and to which a curve-
tion, the largest relative difference between the approximate fitting procedure may be applied.
solution and the exact solution is 1.5 3 10
3 . Taking m ¼
1, the best approximation is R. A. Day and P. H. Morris, The University of Queensland
The author has shown that simple linear regression based
rffiffiffiffiffiffiffiffi
4Tv on equation (8) gives unreliable results. Reasonable results
U¼ 0 < Tv < 0:125 (21a) were obtained only when data points at both small and large
ð time values were discarded. An explanation for this is given
 
8 ð2 1 9ð2 below.
U ¼ 1
2 e
4 Tv þ e
4 Tv Tv . 0:125 (21b)
ð 9 The author also presents a least squares approach for
analysing consolidation test data that involves calculation of
for which U  39.9% at the transition point and the largest complex summations and concludes with the numerical
relative difference is 3.6 3 10
5 . Establishing the transition solution of equation (18). The discussion below demonstrates
from the parabolic approximation to the truncated series for how common numerical tools can be used to simply and
higher values of m is straightforward, but the approximation directly solve the least squares minimisation problem.
in equation (21) is sufficiently accurate for all practical
purposes, and that in equation (20) sufficiently accurate for
most purposes. SIMPLE LINEAR REGRESSION
Problems can arise in least squares adjustments when the Consider first the data at large time values—the late data.
fitting function is discontinuous. While both approximations The author’s example displays significant secondary settle-
(equations (20) and (21)) are continuous in value throughout, ment. Equation (8) does not account for secondary consoli-
to within the numerical accuracy of the calculation, they do dation, and thus the late data must be discarded from the
have a discontinuity in gradient at the transition point. For m regression analysis. Based on conventional methods, the
¼ 0, equation (20), this discontinuity in gradient is approxi- secondary consolidation starts when the reading is about
mately 1.7% of the gradient at the transition point Tv ¼ 3.0 mm (point 13) (Craig, 1995). This is consistent with the
0.213. For m ¼ 1, equation (21), the discontinuity in gradient findings of the author that better results were obtained if
is approximately 0.01% of the gradient at the transition point points 15–17 were excluded.
Tv ¼ 0.125. From our experience of least squares adjustment Consider now the linear regression using the transformed
to fit Fox’s approximation with real consolidation data this data points 1/t 3 and 1=r6t . Fig. 5 shows the result of
discontinuity in the gradient in the fitting function does not regression analysis using points 2–10 plotted on natural
 DISCUSSION 75
5
2·5 3 10 second point plotted (data point No. 3). The first plotted data
point is uniquely distant from the mean. Consequently this
2·0 3 105
first point is disproportionately influential in determining the
intercept. Discarding early time data points (from the right-
hand side) does not eliminate this feature of the data; a
1·5 3 105 similar picture emerges in each case as points are succes-
sively discarded. The average of the remaining points is
always near to the second point from the right because the
1/r 6t

1·0 3 105 1/t 3 values are distributed approximately uniformly on a

logarithmic scale. Fig. 7 shows the different linear regression

5·0 3 104
lines obtained as data points are successively discarded. It
can be seen that the predicted intercept switches from nega-
Primary compression data
tive (impossible) to positive (but too high) values. A reason-
1 Regression line ably accurate estimate is not obtained until only points 7–10
Mean of data points are used in the analysis.
99% confidence interval for regression line As the predicted value of the intercept is used to calculate
25·0 3 104 the value of cv as well as the final consolidation settlement,
0 10 20 30 40 50 60 70
an accurate evaluation of the intercept is very important for
3
1/t the consolidation problem. However, it cannot be obtained
reliably using the simple linear regression technique de-
Fig. 5. Result of linear regression using points 2–10 (natural scribed owing to the nature of the consolidation test data,
scales) even though it is a good predictor of the trend line. The
method is thus intrinsically unsuitable for the analysis of the
scales. Points at larger time values were discarded to avoid consolidation test.
the issue of secondary consolidation. Note that points 5–10
are so close together they appear as one on this plot. Also
shown in Fig. 5 is the 99% confidence interval for the
regression line determined using ordinary statistical methods LEAST SQUARES FITTING
and assumptions. This means there is a 1% probability that The least squares approach presented by the author in-
the true linear relationship does not lie within the bounds volves calculation of complex summations, and concludes
shown. The fit is very good: the correlation coefficient is with the numerical solution of equation (18). A simpler,
0.9998. However, in this problem we wish to determine the transparent and elegant approach is to solve the least squares
intercept, or value of the line at 1/t 3 ¼ 0. The 99% con- problem directly using numerical techniques commonly
fidence interval for the intercept (1=r61 ) is from
1330 to available within spreadsheet programs. The problem can be
1940. Neglecting negative values, which are physically im- stated as: Minimise the objective function Ö, which is
possible, this result means there is 99% confidence that the defined as
value of r1 lies between infinity and 0.283 mm. Simple X X
visual inspection of the settlement–time graph gives a much Ö¼ å2i ¼ w i ðr ti
r t Þ2 (22)
i i
superior result.
Replotting Fig. 5 using logarithmic scales (Fig. 6) enables The weights w i take the value of either 1 or 0 to include or
closer examination of the regression line as 1/t 3 approaches exclude (respectively) the corresponding data point from the
0. It is clear that the regression line neither fits the nearby objective function. The first data point (t ¼ 0) and any data
data points well, nor predicts the intercept accurately. points for which r ti . r1 (representing secondary consoli-
It is a feature of simple linear regression that the mean dation) must be excluded.
values of 1/t 3 and 1=r6t lie exactly on the line. This mean The direct numerical approach does not restrict the theor-
point is also plotted in Figs 5 and 6. The mean value lies etical relationship to a simple approximation or single equa-
towards the right-hand end of the data set, very close to the tion. The classical Fourier series solution (equation (4)) is

106 Primary compression data

Regression line 106
105 Mean of data points Primary compression data
99% confidence interval for regression line Points 2–10
104 Points 3–10
104 Points 4–10
103 Points 5–10
Points 7–10

6

102
1/r t

102
1/r6t

10

1
1
1021

1022
1026 1024 1022 1 102 1022
1/t 3 1026 1024 1022 1 102
1/t 3
Fig. 6. Result of linear regression using points 2–10 (logarith-
mic scales) Fig. 7. Results of linear regression using transformed data
76 DISCUSSION
easily used. All three unknown parameters (r i , r1 , and cv ) McKinley. As long as no data point is near the transition,
can be determined without more effort or difficulty than the issue of discontinuity is not of concern. Even if there is
finding only two parameters. Fig. 8 demonstrates graphically a data point at the transition, using one slope or the other
the good quality of fit that can be obtained using this type does not make any significant difference. For the 72 sets of
of procedure, performed in this case using a common data analysed, this discontinuity does not seem to cause any
spreadsheet program (Day, 2005), for the analysis of a major problem.
consolidation test As for the problems of derivation of parameters from the
The least squares approach is simple yet very powerful, Terzaghi consolidation, this may arise from the change of
and can be easily implemented. It is completely general, and stiffness during the consolidation process and significant
applicable to an enormous range of problems. Multi-variable secondary consolidation settlement. El-Gehani (2004) has
problems and complex relationships requiring multi-stage tried an automatic procedure that attempted to remove points
calculations can be fitted as easily as simple ones. Disconti- with significant secondary consolidation. It met with a good
nuities do not present problems. Multiple relationships can degree of success for the data set from Craig (1995), but as
be used to cover the full range of data. For example, the 72 sets of data used do not show significant secondary
equation (23) (Craig, 1995) is very simple to implement for consolidation behaviour, the result cannot be seen as con-
the consolidation test. clusive. Furthermore, for the 72 data sets tried, the resulting
ð 2 curve fits the data points reasonably well: therefore it can be
For U < 0:6, Tv ¼ U concluded that the variation of stiffness may not be signifi-
4 cant for the 72 sets of data used.
For U . 0:6, Tv ¼
0:933 log ð1
U Þ
0:085 (23) The author would also like to thank Day & Morris for
their detailed study on the simple linear regression scheme
that the author tried but failed to investigate further. The
Author’s reply
author is glad to find that the conclusion from the detailed
First the author would like to thank the two groups of
study is that ‘the method is thus intrinsically unsuitable for
colleagues who have spent considerable time and effort to
the analysis of the consolidation test’.
investigate and discuss the technical note concerned.
We also tried using a direct minimisation using the same
Doran & McKinley were right in saying that the author is
equation as equation (22). It works for some data sets but
unaware of the paper by Fox (1948), but equations (19a) and
not all the data sets tried. We found that the direct minimum
(19b) are the same as equations (1) and (6). From Table 1,
could be very flat, and Microsoft Excel Solver failed to
the U values using equation (1) and equation (6) for Tv
achieve the absolute minimum. However, we discovered an
values of 0.2 and 0.3 are 0.5046 and 0.6180, 0.5051 and
alternative minimum that is much more suitable for Excel
0.6134 respectively, whereas the exact solutions are 0.5041
calculation. Instead of solving equation (18), the author
and 0.6132. The author feels that they are close enough for
realised that it is equivalent to finding the maximum of the
pffiffiffiffiffiffiffiffiffiffiffi
all practical purposes in geotechnical application.
function A(cv )= B(cv ), which is more stable than the solu-
The author agrees that Fox’s approximation, whether the
tion of equation (18) for Microsoft Excel.
transition point is at U ¼ 0.5 or 0.6, is sufficiently accurate
In El-Gehani (2004) we also tried the three-variable
for any likely application, and is also significantly more
method, which uses the initial settlement, the final primary
accurate than the Hansen approximation used in the techni-
settlement and the coefficient of consolidation as variables.
cal note. The same conclusion was drawn by the author even
The scheme was devised before the receipt of the discussion
before the technical note was published. Therefore the equa-
from Day & Morris, and it is good to know that they agree
tion pair (1) and (6) with the transition at U ¼ 0.6 has
that this is a good way forward. However, from our experi-
formed the basis of Tan (2003). Instead of using the
ence, the initial settlements obtained are very sensitive to
‘artificial’ example from Craig (1995), 72 sets of real
the quality of data, and can be very different from the
consolidation data have been obtained from three different
traditional method. The automatic procedure used in El-
industrial and academic sources. The results obtained are in
Gehani (2004) was a similar procedure to equation (22).
good agreement, especially with Taylor’s method, for ob-
In conclusion, the author is very grateful for the general
vious reasons.
interest that the technical note has generated. It is good to
The author also agrees that Fox’s approximation or the
know that investigations carried out in three different coun-
equation pair (1) and (6) with the transition at U ¼ 0.6 is
tries came to similar approaches, with conclusions support-
much more convenient to use than the series solution given
ing each other. It is his sincere hope that further
by equation (4). Concerning the discontinuity of slope at the
development could be made in the automatic determination
interface, the author has the same experience as Doran &
of the coefficient of consolidation, thus removing operator
bias and reducing the time and effort required in oedometer
Time: min tests, while highlighting soil that has unusual consolidation
0·1 1 10 100 1000 10000 characteristics.
0
Data points
Fitted relationship
REFERENCES
Settlement, r: mm

1 Day, R. A. (2005). http://www.uq.edu.au/robertday/consolidation.

Accessed 25 August 2005
El-Gehani, G. A. (2004). New method for the determination of the
coefficient of consolidation using a fully automatic least squares
2 procedure. MSc thesis, University of Birmingham.
Fox, E. N. (1948). The mathematical solution for the early stages of
consolidation. Proc. 2nd Int. Conf. Soil Mech. Found. Engng,
Rotterdam 1, 41–42.
3 Tan, S. N. (2003). New method for the determination of the coeffi-
cient of consolidation, cv from the oedometer results. MSc
Fig. 8. Comparison of data points and fitted relationship thesis, University of Birmingham.