Posted in Geotechnical Engineering, STADYN, TAMWAVE

Relating Hyperbolic and Elasto-Plastic Soil Stress-Strain Models

It is routine in soil mechanics to attempt to use continuum mechanics/theory of elasticity methods to analyse the stresses and strains/deflections in soil.  We always do this with the caveat that soils are really not linear in their response to stress, be that stress axial, shear or a combination of the two.  In the course of the STADYN project, that fact became apparent when attempting to establish the soil modulus of elasticity.  It is easy to find “typical” values of the modulus of elasticity; applying them to a given situation is another matter altogether.  In this post we will examine this problem from a more theoretical/mathematical side, but one that should vividly illustrate the pitfalls of establishing values of the modulus of elasticity for soils.

Although the non-linear response of soils can be modelled in a number of ways, probably the most accepted method of doing so is to use a hyperbolic model of soil response.  This is illustrated (with an elasto-plastic response superimposed in red) below.

Hyberbolic-Stress-Strain

The difficulties of relating the two curves is apparent.  The value E1 is referred to as the “tangent” or “small-strain” modulus of elasticity.  (In this diagram axial modulus is shown; similar curves can be constructed for shear modulus G as well.)  This is commonly used for geophysical methods and in seismic analyses.

As strain/deflection increases, the slope of the curve decreases continuously, and the tangent modulus of elasticity thus varies continuously with deflection.  For larger deflections we frequently resort to a “secant” modulus of elasticity, where we basically draw a line between the origin (usually) and whatever point of strain/deflection we are interested in.

Unfortunately, like its tangent counterpart, the secant modulus varies too.  The question now arises: what stress/strain point do we stop at to determine a secant modulus?  Probably a better question to ask is this: how do we construct an elasto-plastic curve that best fits the hyperbolic one?

One solution mentioned in the original study is that of Nath (1990), who used a hardening model instead of an elastic-purely plastic model.  The difference between the two is illustrated below.

Elasto-Plastic Response

Although this has some merit, the elastic-purely plastic model is well entrenched in the literature.  Moreover the asymptotic nature of the hyperbolic model makes such a correspondence “natural.”

Let us begin by making some changes in variables.  Referring to the first figure,

y=\frac{\sigma}{\sigma_1-\sigma_3}=\frac{\sigma}{\sigma_0}

and

x = \epsilon

Let us also define a few ratios, thus:

A_1 = \frac {E_1}{\sigma_0}

A_2 = \frac {E_2}{\sigma_0}

A = \frac {E_2}{E_1}

Substituting these into the hyperbolic equation shown above, and doing some algebra, yields

y=\frac{x A_1}{1+x A_1}

One way of making the two models “close” to each other is to use a least-squares (2-norm) difference, or at least minimising the 1-norm difference.  To do the latter with equally spaced data points is essentially to minimise the difference (or equate if possible) the integrals of the two, which also equates the strain energy.  This is the approach we will take here.

It is easier to equate the areas between the two curves and the \sigma_0 line than to the x-axis.  To do this we need first to rewrite the previous equation as

y'=1-\frac{x A_1}{1+x A_1}

Integrating this with respect to x from 0 to some value x_1 yields

A_{hyp} = \frac{ln\left( 1+x_1 A_1 \right)}{A_1}

Turning to the elastic-plastic model, the area between this “curve” and the maximum stress is simply the triangle area above the elastic region.  Noting that

E_2 = \frac {\sigma_0} {x_2} ,

employing the dimensionless variables defined above and doing some additional algebra yields the area between the elastic line and the maximum stress, which is

A_{ep} = \frac {1}{2 A A_1}

Equation the two areas, we have

ln\left( 1+x_1 A_1 \right) - \frac {1}{2 A} = 0

With this equation, we have good news and bad news.

The good news is that we can (or at least think we can) solve explicitly for A , the ratio between the elastic modulus needed by elasto-plastic theory and the small-deflections modulus from the hyperbolic model.  The bad news is that we need to know A_1 , which is the ratio of the small deflections modulus to the limiting stress.  This implies that the limiting stress will be a factor in our ultimate result.  Even worse is that x_1 is an input variable, which means that the result will depend upon how far we go with the deflection.

This last point makes sense if we consider the two integrals.  The integral for the elasto-plastic model is bounded; that for the hyperbolic model is not because the stress predicted by the hyperbolic model is asymptotic to the limiting stress, i.e., it never reaches it.  This is a key difference between the two models and illustrates the limitations of both.

Some additional simplification of the equation is possible, however, if we make the substitution

x_1 = n x_2

In this case we make the maximum strain/deflection a multiple of the elastic limit strain/deformation of the elasto-plastic model.  Since

x_2 = \frac {sigma_0}{E_2} = \frac {1}{A_2} = \frac {1}{A A_1}

we can substitute to yield

ln\left( 1+\frac{n}{A} \right) - \frac{1}{2A} = 0

At this point we have a useful expression which is only a function of n and A .  The explicit solution to this is difficult; the easier way to do this is numerically.  In this case we skipped making an explicit derivative and use regula falsi to solve for the roots for various cases of n .  Although this method is slow, the computational time is really trivial, even for many different values of n .  The larger value of n , the more deflection we are expecting in the system.

The results of this survey are shown in the graph below.

hyperb

The lowest values we obtained results for were about n = \frac{x_1}{x_2} = 0.75 .  When n = \frac{x_1}{x_2} = 1 , it is the case when the anticipated deflection is approximately equal to the “yield point.”  For this case the ratio between the elasto-plastic modulus and the small-strain hyperbolic modulus is approximately 0.4.  As one would expect, as n increases the elasto-plastic system becomes “softer” and the ratio A = \frac {E_2}{E_1} likewise decreases.  However, as the deflection increases this ratio’s increase is not as great.

To use an illustration, consider pile toe resistance in a typical wave equation analysis.  Consider a pile where the quake (x_2 ) is 0.1″.  Most “traditional” wave equation programs estimate the permanent set per blow to be the maximum movement of the pile toe less the quake.  In the case of 120 BPF–a typical refusal–the set is 0.1″, which when added to the quake yields a total deflection of 0.2″ of a value of n = 2.  This implies a value of A = 0.2139950 .  On the other hand, for 60 BPF, the permanent set is 0.2″, the total movement is 0.3″, and n = 3, which implies a value of A = 0.1713409.  Cutting the blow count in half again to 30 BPF yields n = 5 or A = 0.1383195.  Thus, during driving, not only does the plastic deformation increase, the effective stiffness of the toe likewise decreases as well.

Based on all this, we can draw the following conclusions:

  1. The ratio between the equivalent elasto-plastic modulus and the small-strain modulus decreases with increasing deflection, as we would expect.
  2. As deflections increase, the effect on the the equivalent modulus decreases.
  3. Any attempt to estimate the shear or elastic modulus of soils must take into consideration the amount of plastic deformation anticipated during loading.  Use of “typical” values must be tempered by the actual application in question; such values cannot be accepted blindly.
  4. The equivalence here is with hyperbolic soil models.  Although the hyperbolic soil model is probably the most accurate model currently in use, it is not universal with all soils.  Some soils exhibit a more definite “yield” point than others; this should be taken into consideration.
Advertisements
Posted in Deep Foundations, STADYN

Shaft Friction for Driven Piles in Clay: Alpha or Beta Methods?

In a previous post we discussed beta methods for driven pile shaft friction in sands, which are pretty much accepted, although (as always) the values for \beta can vary from one formulation to the next.  With clays, also as always, things are more complicated.

Since the researches of Tomlinson in the 1950’s, the shaft friction of piles in clays has been thought to be a function of the undrained shear strength of the clay multiplied by an adhesion factor \alpha , thus

f_s = \alpha c_u

This was seriously challenged by Burland (1973) who noted the following:

Whereas the use of undrained shear strength for calculating the end bearing capacity of a pile appears justified there seems little fundamental justification for relating shaft adhesion to undrained strength for the following reasons:

  1. the major shear distortion is confined to a relatively thin zone around the pile shaft (Cooke and Price (1973)).  Drainage either to or from this narrow zone will therefore take place rapidly during loading;

  2. the installation of a pile, whether driven or cast-in situ, inevitably must disturb and remould the ground adjacent to the pile shaft;

  3. quite apart from the disturbance caused by the pile there is no simple relationship between the undrained strength and drained strength of the ground.

Burland buttressed his case by noting that

\beta = K tan \phi

and presenting a graph similar to the following:

Beta Image 1

where, as seen earlier,

  • K_o = 1 - sin \phi is in red.
  • tan \phi is in blue.
  • \beta is in green.

Since, for the ranges of drained friction angles for clay (20-25 deg.) the value for \beta was relatively constant, value of \beta were relatively invariant with friction angle, and thus could be estimated with relative accuracy.  His empirical correlation was very successful with soft clays, not as much with stiff ones.

The year after Burland made his proposal, McClelland (1974) noted the following:

It is not surprising that there is a growing dissatisfaction with attempts to solve this problem through correlations of \alpha with c_u .  This is accompanied by a growing conviction that pile support in clay is frictional in character–that load transfer is dependent upon the effective lateral pressure acting against the side of the pile after it is driven.

However, \beta methods–which would embody McClelland’s preferred idea–have never been universally accepted for pile shaft friction in clays.  A large part of the problem, as noted by Randolph, Carter and Wroth (1979) is that the lateral pressure itself is dependent upon the undrained shear strength of the soils.  It is thus impossible to completely discount the effect of undrained shear strength on the shaft friction, even with the remoulding Burland and others have noted.

This has led to the “hybrid” approach of considering both undrained shear strength and effective stress.  This is embodied in the American Petroleum Institute (2002) specification.  A more advanced version of this is given in Kolk and van der Velde (1996).  They give the \alpha factor as

\alpha = 0.9\left( \frac {L-z} {d} \right)^{-0.2} \left( \frac {c_u} {\sigma'_{vo}} \right)^{-0.3} \leq 1

The notation is the same as in this post except that we add c_u , which is the undrained shear strength.

In this case the unit shaft friction is given by the equation

f_s = 0.9\left( \frac {L-z} {d} \right)^{-0.2} \left( \frac {c_u} {\sigma'_{vo}} \right)^{-0.3} c_u

There are a couple of things worth noting about this.

The first is that we can transform this into a \beta method of the form

f_s = \beta \sigma'_{vo}

with the following multiplication

f_s = 0.9\left( \frac {L-z} {d} \right)^{-0.2} \left( \frac {c_u} {\sigma'_{vo}} \right)^{0.7} \sigma'_{vo}

(A similar operation appears in Randolph (2005).)

in which case

\beta = 0.9\left( \frac {L-z} {d} \right)^{-0.2} \left( \frac {c_u} {\sigma'_{vo}} \right)^{0.7}

The only thing we would have to do is to find a way to incorporate the limiting condition for \alpha , which we will discuss shortly.

The second thing is that the term \left( \frac {L-z} {d} \right) appears in both this formulation and that for sands in this post.  The difference is that, while Kolk and van der Velde (1996) use the term in a power relationship, Randolph (2005) uses it in an exponential way.  The basic concept in both is the same: the term is at a maximum at the pile toe and decays toward the mudline.

The two are compared in the figure below.

kandvdv-vs-randolph

Here the quantity \left( \frac {L-z} {d} \right) is at the x-axis and the following is at the y-axis:

  • Kolk and van der Velde Method for Clays, \left( \frac {L-z} {d} \right)^{-0.2} in red.
  • Randolph Method for Sands, e^{-\mu \left( \frac {L-z} {d} \right)} in blue, where \mu = 0.05 .
  • e^{-\mu \left( \frac {L-z} {d} \right)} in green, where \mu = 0.02 .

The graph illustrates the problem (from a computational standpoint) with the Kolk and van der Velde method: there is a singularity in their coefficient using the power relationship at the pile toe, while the exponential relationship yields a value of unity at this point.  The last correlation in green is approximately the best fit of the exponential relationship with the power relationship of Kolk and van der Velde, using either 1-norm or 2-norm methods.  It is not very good; it would be interesting, however, to see what kind of value for \mu might result if this had been in Kolk and van der Velde’s original statistical correlation equation.

In view of all this, perhaps the best way to enforce the limit is to do so as follows:

\left( \frac {L-z} {d} \right)\geq1

From all this, we can say that it is certainly possible to compute shaft friction for driven piles with a \beta method provided we include the effects of the undrained shear strength.

References

In addition to the original study and previous posts, the following references are noted:

Kolk, A.J., and van der Velde, A. (1996) “A Reliable Method to Determine Friction
Capacity of Piles Driven into Clays.” Proceedings of the 28th Offshore Technology Conference, Houston, TX, 6-9 May.  OTC 7993.

McClelland, B. (1974) “Design of Deep Penetration Piles for Ocean Structures.”  Journal of the Geotechnical Engineering Division, ASCE, Vol. 111, July.

Posted in Geotechnical Engineering, STADYN

Lateral Earth Pressure Coefficients for Beta Methods in Sands

In our last post we considered some basic concepts behind beta methods for determining beta coefficients for estimating shaft friction for piles in sands.  The idea is that the unit friction along the surface of the pile can be determined at any point by the relationship

f_s = \beta \sigma'_{vo}

where f_s is the unit shaft friction, \sigma'_{vo} is the vertical effective stress, and \beta is the ratio of the two, which can be further broken down as follows:

\beta = K tan \phi

where K is the lateral earth pressure coefficient and \phi is the internal friction angle of the soil.  Our last post showed that, when compared with empirically determined values of \beta , values of K determined from more conventional retaining wall considerations are not adequate to describe the interaction between the shaft of the pile and the soil.

Needless to say, there has been a good deal of research to refine our understanding of this relationship.  Also, needless to say, there is more than one way to express this relationship.  The formulation we will use here is that of Randolph, Dolwin and Beck (1994) and Randolph (2003), and was recently featured in Han, Salgado, Prezzi and Zaheer (2016).  The basic form of the lateral earth pressure equation is as follows:

K = K_{min} + (K_{max} - K_{min}) e^{-\mu \frac {L-z}{d}}

Let’s start on the right end of the equation; the exponential term is a way of representing the fact that the maximum shaft friction (with effective stress taken into account) is just above the pile toe and decays above that point to the surface of the soil.  This was first proposed by Edward Heerema (whose company was instrumental in the development of large steam and hydraulic impact hammers) in the early 1980’s.  (For another paper of his relating to the topic, click here.)

In any case the variables in the exponential term are as follows:

  • \mu = rate of exponential decay, typically 0.05
  • L = embedded length of pile into the soil
  • z = distance from soil surface to a given point along the pile shaft.  At the pile toe, L = z and L-z = 0 , and the exponential term becomes unity.
  • d = “diameter” of the pile, more commonly designated as B in American textbooks.

K_{min} is the minimum lateral earth pressure coefficient.  It, according to Randolph, Dolwin and Beck (1994) “can be linked to the active earth pressure coefficient.” Randolph (2003) states that its value lies in the range 0.2-0.4. We stated in our previous post that

K_a = \frac {1 - sin \phi} {1 + sin \phi}

How do these two relate?  Although in the last post we produced extensive parametric studies on these, a simpler representation is to compare the active earth pressure coefficient with Jaky’s at-rest coefficient, which is done below.

Beta Image 4

The at-rest coefficient from Jaky is in blue and the active coefficient from Rankine is in red.  The range of 0.2 < K_a < 0.4 approximately translates into 25^\circ < \phi < 45^\circ , which is a wide range for granular soils but reasonable.

That leaves us K_{max}.  Randolph, Dolwin and Beck (1994) state that

K_{max} = S_t N_q

N_q , of course, is the bearing capacity factor at the toe.  It may seem odd to include a toe bearing capacity factor in a shaft equation, but keep in mind that cavity expansion during pile installation begins (literally) with an advancing toe.  Typically 8 < N_q < 40 depending upon whether the sand is loose (low end) or dense (high end.)  S_t “is the ratio of the radial effective stress acting in the vicinity of the pile tip at shaft failure to the end-bearing capacity.”  Values for S_t vary somewhat but generally centre around 0.02.  This in turn implies that 0.16 < K_{max} < 0.8 .  Inspection of the complete equation for K shows that, if L = z and the exponential term is at its maximum, K_{min} cancels out and the range of K_{max} is a range for K .

Comparing this result to the graph above, for larger values of \phi these values of K are greater than those given by Jaky’s Equation, which is what we were looking for to start with.  To compute \beta , we obviously will need to multiply this by tan \phi (or tan \delta ).  For, say, \delta = 35^\circ , this leads to \beta_{max} = 0.8 \times tan 35^\circ = 0.56 .  By way of comparison, using Jaky’s Equation for K, \beta = (1 - sin 35^\circ) tan 35^\circ = 0.30 .

From this we have “broken out” of Burland’s (1973) limitation on \beta , which was useful for him (and will be useful to us) for some soils but creates problems with higher values of \phi   Although some empirical methods indicate higher values for \beta , if we consider variations in S_t and other factors, this differential can be minimised, and in any case this is not a rigourous excercise but a qualitative one.

One thing we should further note–and this is important as we move forward–is that there is more than one way to compute K_{max} .  Randolph (2003) states that, when CPT data is available, it can be computed as follows for open-ended piles:

K_{max} = 0.01 \frac {q_c}{\sigma'_{vo}}

where q_c is the cone tip resistance.  Randolph (2003) recommends the coefficient be increased to 0.015 for closed-ended piles.  Making generalisations from this formulation is more difficult than the other, but the possibility of using this in conjunction with field data is attractive indeed.

At this point we have a reasonable method of computing \beta coefficients.  However, we still have the issue of clay soils to deal with, and this will be done in a subsequent post.

References

In addition to those previously given, we add the following:

  • Han, F., Prezzi, M., Salgado, R. and Zaheer, M., (2016), “Axial Resistance of Closed-Ended Steel-Pipe Piles Driven in Multilayered Soil“, Journal of Geotechnical and Geoenvironmental Engineering, DOI: 10.1061/(ASCE)GT.1943-5606.0001589.
  • Randolph, M., Dolwin, J., Beck, R. 1994, ‘Design of driven piles in sand’, GEOTECHNIQUE, 44, 3, pp. 427-448.
  • Randolph, M. 2003, ‘Science and empiricism in pile foundation design’, GEOTECHNIQUE, 53, 10, pp. 847-875.

 

Posted in Geotechnical Engineering, STADYN

A First Look at Estimating Beta Factors for Determining Pile Shaft Capacity of Driven Piles

In the last posting about STADYN, we put forth considerations for interface elements between the pile shaft and the soil.  Before we formally incorporate these into the model (or whether we will incorporate them or not) some consideration of how the interface actually works.  We will start those considerations by looking at methods by which the static capacity of driven piles is computed, and specifically the so-called “beta” methods which are used for cohesionless and sometimes cohesive soils.

Beta methods assume that the shaft resistance of the pile is a function of the effective stress of the soil along the pile shaft.  They assume that the horizontal stress that results from the vertical stress acts perpendicular to the surface of the pile.  The pile surface thus acts like a block on a surface with some kind of Coulombic friction acting against the downward settlement of the pile.  The beta coefficient is the ratio between the vertical effective stress and the horizontal friction on the pile, or

f_s = \beta \sigma'_o

\beta is in turn broken down into two components: the lateral earth pressure coefficient, which is the ratio between the horizontal and vertical stresses,

K = \frac {\sigma_h}{\sigma'_o}

and the coefficient of friction, or

\mu = tan \phi

We put these together to yield

\beta = K tan \phi

At this point let’s make two assumptions.  The first is that the lateral earth pressure coefficient is in fact the at-rest lateral earth pressure coefficient.  (For some discussion of this, you can view this slide presentation.)  The second is that the friction angle between the pile and the soil is in fact the same as the soil’s internal friction angle.  If we use Jaky’s formula for the at-rest condition, these assumptions yield

\beta = \left (1-\sin(\phi)\right )\tan\phi

The various components of this equation are plotted below.

Beta Image 1

The three lines are as follows:

  • K_o = 1 - sin \phi is in red.
  • tan \phi is in blue.
  • \beta is in green.

It’s interesting to note that, as K_o increases, tan \phi decreases, and so \beta is within a surprisingly narrow range of values.  This plot is similar to one shown in Burland (1973), which we will discuss later.

If this were the case in practice, estimating \beta would be a straightforward proposition.  We’ll take two examples to show that this is not the case.

Let’s start with the Dennis and Olson Method for cohesionless soils, which is described here.  To arrive at \beta they do the following:

  1. They add a depth factor, which we will not consider.  Depth factors and critical lengths are common in static methods, but they are not well documented in the field.
  2. They assume K_o = 0.8 if their values for friction angle are used.
  3. They vary the friction angle from 15-35 degrees depending upon the type of soil.

Leaving out the depth factor, for this method \beta ranges from 0.21 to 0.56.  This is a considerably wider variation than is indicated above.  Since the depth factor is frequently greater than unity, this range is even larger.

An easier way to see this is to consider the method of Fellenius.  His values for \beta are as follows:

  • 0.15-0.35 for clay
  • 0.25-0.50 for silt
  • 0.30-0.90 for sand
  • 0.35-0.80 for gravel

Again the range of values is greater than the figure above would indicate.  Why is this?

Although it’s tempting to use a straight empirical approach, let’s back up and consider the structure of the basic equation about and the assumptions behind it.  There are several ways we can alter these equations in an attempt to match field conditions better by considering these assumptions and seeing what changes might be made.

The Two Friction Angles Aren’t the Same

The first one is suggested by the notation in Dennis and Olson: the internal friction angle of the soil and that of the soil-pile interface are not the same.  Retaining wall theory (when it considers friction) routinely makes this assumption; in fact, the ratio \frac {\delta}{\phi} routinely appears in calculations.  Let us rewrite the equation for \beta as

\beta = (1 - sin \phi) tan \delta

and be defining the ratio

m = \frac {\delta}{\phi}

we have

\beta = (1 - sin \phi) tan (m \phi)

If we plot this in a three-dimensional way, we get the following result.

Beta Image 2.png

\beta is the vertical axis; m is varied from 0.25 to 1.75.  The results show that, for a given \phi , if we increase m we will increase \beta , and this increase is much more pronounced at higher values of \phi .

Although it’s certainly possible to have very high values of \delta = m \phi , as a practical matter in most cases m < 1.  Nordlund’s Method, for example indicates that m > 1 only with tapered piles, where a tapered pile face induces some compression in the soil in addition to shear.  In any case is m < 1 this will tend to depress values of \beta .  We should also note that using a ratio m does not mean that it will be a constant for any given soil.  This is especially true if \phi = 0 , where a multiplier is meaningless and we should have recourse to an additive term as well.

Jaky’s Equation Doesn’t Apply, or At-Rest Earth Pressure Conditions Are Not Present

Another assumption that can be challenged is that Jaky’s Equation doesn’t apply, or we don’t have at-rest earth pressure conditions.  Although Jaky’s Equation has done well, it is certainly not the last word on the subject, especially for overconsolidated soils (which we will discuss below.)  To try to “cover our bases” on this, let’s consider a range of lateral earth pressure coefficients by assuming that Jaky’s Equation is valid for the at-rest condition and that we need to somehow vary between some kind of active state and passive state.  The simplest way to do this is to assume Rankine’s conditions with level backfill, which just happens to be identical to Mohr-Coulomb relationships between confining and driving stresses.  (OK, it’s not all luck here…)  Thus,

K_a = \frac {1-sin\phi}{1+sin\phi}

and

K_p = \frac {1+sin\phi}{1-sin\phi}

Let us also define an active-passive factor called actpas, where actpas = -1 for the active state, 0 for the at-rest state and 1 for the passive state.  We then plot this equation

\beta = K(\phi,actpas) tan \phi

below.  Since we only have K values for three values of actpas, we’ll use a little Lagrangian interpolation in an attempt to achieve a smooth transition between the states.

Beta Image 3

We note from this the following:

  1. The dip in \beta for the high values of \phi and  -1 < actpas < 0 (states tending towards the active) may be more a function of the interpolation than the physics.  OTOH, if we look at NAVFAC DM 7.02, Chapter 3, Figure 1, we see a dip between the at-rest and active states for dense sands, which is what we would expect at higher values of \phi .
  2. Values of \beta for the active case show little variation.  Given that driven piles are subject to cavity expansion during installation, one would expect some passivity in earth pressures.  Drilled shafts are another story; however, if we look, for example, at O’Neill and Reese, values for \beta can certainly range higher than one sees with the active states above.  Bored piles, however, are beyond the scope of this discussion.
  3. For low values of \phi , there is little variation between the three states.
  4. If we compare these values with, say, those of Fellenius or Dennis and Olson, we cannot say that the fully passive state applies for most reasonable values of \phi , undrained or drained.  (Values in Nordlund, however, indicate higher values of K for larger displacements, approaching full passivity for large displacement piles.  Another look at this issue is here.)

Conclusion

If we compare the results we obtain above with empirical methods for determining \beta , we see that none of the variations shown above really allows us to match the theory we’ve presented with the empirical methods we’ve described (and others as well.)  As a general rule, \delta < \phi or m < 1 , so it’s safe to conclude that our assumption that the K can be determined using Jaky’s Equation only results in values of \beta that are too low.

It’s tempting to simply fall back on an empirical value for \beta , but for finite element analysis a more refined approach seems appropriate.  In subsequent posts we’ll look at such an approach, along with the issue of applying \beta methods to cohesive soils as well as cohesionless ones.

References

In addition to those in the original study, the following reference is mentioned here:

  • Burland, J.B. (1973) “Shaft friction of piles in clay – A simple fundamental approach.” Ground Engineering 6(3):30-42, January.
Posted in STADYN

STADYN Wave Equation Program 3: Match Quality vs. Least Squares Analysis

Having broached the subject of Poisson’s Ratio and how it is computed for forward methods, we can turn to how it affects inverse methods.  However, at the same we need to consider an issue that is vital to understanding either this method or methods such as CAPWAP: how the actual pile head signal is matched with the signal the model proposes.  There is more than one method of doing this, and the method currently used by CAPWAP is different than what is widely used in many engineering applications.  Is this difference justified?  First, we need to consider just what we are talking about here, and to do that we need a brief explanation of vector norms.

Vector Norms

A vector is simply a column (or row) of numbers.  We want to compare vectors in a convenient way.  To do this we must aggregate the entries in the vector into a scalar number, and we use what we call norms to accomplish this.  In theory there are an infinite number of ways to do this: according to this reference, there are three types of norms in most common use, they are as follows:

stadyn3norms
Left: the 1-norm.  Center: the 2-norm, or Euclidean norm.  Right: the infinity-norm.

So how do use norms in signal matching? We reduce the force-time (or in our case the velocity-time) history at the pile top after impact into a series of data points, and then for each point of time of each data point we compute the results our proposed model gives us and subtract it from the actual result.  In the equation above each data point is a value x_j .  When we have all the differences in hand, we take them and compute a vector of differences, and then in turn take the norm of those differences.  We do this successively by changing parameters until we get a norm value which is the minimum we can reach.  For the STADYN program, we use the \xi - \eta values as parameters and iterate using a polytope method (standard or annealed, for our test case the latter.)

For our purposes the infinity norm can be eliminated up front: in addition to having uniqueness issues (see Santamarina and Fratta (1998), we have enough of those already) it only concerns itself with the single largest difference between the two data sets.  Given the complexities of the signal, this is probably not a good norm to use.

That leaves us with the 1-norm and 2-norm.  To keep things from getting too abstract we should identify these differently, as follows:

  1. 1-norm = “Match Quality” for CAPWAP (see Rausche et. al. (2010))
  2. 2-norm = Least Squares or Euclidean norm (think about the hypotenuse of a triangle.)  This relates to many methods in statistics and linear algebra, and has a long history in signal matching (Manley (1944).)  This is what was used in the original study.

One thing that should be noted is that the norm we actually use is modified from the above formulae by division of the number of data points.  This is to prevent mishap in the event the time step (and thus the number of data points) changes.  However, for the Mondello and Killingsworth (2014) pile, the wall thickness of the steel section drove the time step, which did not change with soil changes; thus, this division is immaterial as long as it is done every time, which it was.

Application to Test Case

As noted earlier, we will use the four-layer case using the annealed polytope method of matching.  Let us start at the end, so to speak, by showing the static load test data that the program runs with the final configuration:

Davisson Load, kN Original \nu \phi -based \nu % Change
1-norm 278 300 7.91%
2-norm 187.1 218 16.52%
% Change 48.58% 37.61%

The runs were done for both the original Poisson’s Ratio (\nu ) and that computed using the internal friction angle \phi .  The CAPWAP run done on this project recorded a Davisson simulated static load capacity of 146.3 kN.

Changing the way \nu is computed produces larger variations in SRD (soil resistance to driving) for each of the norms than was evident in the last post.  This is because the values of \nu now vary with \phi , which overall tends to increase the value of \nu for the same value of \phi .  We will discuss this in detail below.

The most dramatic change took place with the norm was changed; the value for SRD is a third to a half higher with the Least Squares solution, depending upon the way \nu is computed.

xi results
Layer Original nu, 1-norm Original nu, 2-norm Phi-based nu, 1-norm Phi-based nu, 2-norm
Shaft Layer 1 -0.708 -0.812 -0.686 0.471
Shaft Layer 2 -0.709 -0.751 -0.845 -0.96
Shaft Layer 3 -0.71 -0.984 0.966 -0.439
Shaft Layer 4 -0.586 0.428 -0.75 0.196
Pile Toe -0.69 0.366 -0.491 0.804
Average -0.681 -0.351 -0.361 0.014

The values of \xi (degree of cohesion) tend to decrease for the Match Quality but the opposite for the Least Squares method.  it is interesting to note that the Least Squares \phi based \nu is the only run to venture into predominately cohesive territory (\xi > 0 , which is interesting in a soil which is generally characterized as cohesive.

eta results
Layer Original nu, 1-norm Original nu, 2-norm Phi-based nu, 1-norm Phi-based nu, 2-norm
Shaft Layer 1 -1.71 -0.622 -8.68 -1.08
Shaft Layer 2 -1.62 -1.38 3.29 -0.117
Shaft Layer 3 -0.838 -4.373 -1.86 -5.85
Shaft Layer 4 -1.74 -28.363 -14 -27.5
Pile Toe -1.29 1.814 8.19 1.52
Average -1.440 -6.585 -2.612 -6.605

The values of \eta (consistency or density) are all low, but more so for the Least Squares cases than the Match Quality cases.  Low values of \eta are to be expected in a soil like this, but these tend to be extreme.  Although limiters such as for elastic modulus are included to prevent serious misadventure in the soil properties, the existence of extreme values of \eta is something that needs to be re-examined.  (-1 < \xi < 1 by physical necessity.)

Poisson’s Ratio Result
Layer Original nu, 1-norm Original nu, 2-norm Phi-based nu, 1-norm Phi-based nu, 2-norm
Shaft Layer 1 0.279 0.269 0.45 0.45
Shaft Layer 2 0.279 0.275 0.158 0.312
Shaft Layer 3 0.279 0.252 0.45 0.45
Shaft Layer 4 0.291 0.393 0.45 0.45
Pile Toe 0.281 0.387 0 0.45
Average 0.282 0.315 0.302 0.422

As was the \eta values, the values of \nu tend to increase with the \phi based values.  The Match Quality \phi -based values are highly irregular, which in turn reflect the wide swings in \eta with less cohesive values of \xi .

Now let us present the optimization tracks for each of these cases.

 

The original study discusses the numbering system for the xi and eta parameters.  In short, tracks 1-6 are for the shaft and 7-8 are for the toe.  From these we can say the following:

  1. The Match Quality runs tend to converge to a solution more quickly. The x-axis is the number of steps to a solution.
  2. The Match Quality run tended to eta values that were more “spread out” while the Least Squares solution tended to have one or two outliers in the group.
  3. The runs go on too long.  This is because, in the interest of getting a working solution, the priority of stopping the run at a convergence was not high.  This needs to be addressed.

Now the norms themselves should be examined as follows:

Final Norm Original Nu Phi-Based Nu % Change
1-norm 0.148395682775873 0.134369614266467 -9.45%
2-norm 0.001494522212204 0.001456397402301 -2.55%

In both cases the difference norms decreased with the \phi -based nu , the Match Quality difference was more pronounced.  The difference norm for the Match Quality is higher than the Least Squares solution, which is to be expected.

We finally look at the tracks compared with each other for the four cases.

 

It’s tempting to say that the Match Quality results “track more closely” but the whole idea of using a norm such as this is to reduce the subjective part of the analysis.  However, this brings us to look at why one norm or the other is used.

The Least Squares analysis is widely used in analyses such as this.  It is the basis for almost all regression analysis.  However, the Match Quality has some advantages.  It is considered more “robust” in that it is less sensitive to outliers in the data.  In this case, the most significant outlier is the region around L/c = 1.5, which was discussed in the original study.  Situations such as this reveal two sources of uncertainty in the model: the integrity of the mounting of the instrumentation, and the accuracy of the pile data (lengths, sizes, acoustic speed of the wood, etc.) The Match Quality certainly can help to overcome deficiencies caused by this and other factors.  Whether this is at the expense of accuracy has yet to be determined.

So we are left with two questions:

  1. If we were to improve the quality of the data by addressing the present and other issues, would we be better off if we used Least Squares?  The answer is probably yes.  Getting this in the field on a consistent basis is another matter altogether.
  2. Will the two methods yield different results?  With STADYN this is certainly the case; the use of the Match Quality with STADYN however yields results that are double those of CAPWAP.  With CAPWAP we have no way of comparing the two; the Match Quality is all we have.

Conclusions

Based on all of this we conclude the following:

  1. The use of a \phi based \nu leads to an improvement in the signal matching, due probably to the reduction in the number of real parameters being considered.  It will probably remain as the default option.
  2. Any final conclusions on this topic depend upon limiting the values of \eta “within the box” to prevent serious outliers.  This will be the topic of future study.
  3. We also need to address the issue of stopping the runs at a more appropriate point.
  4. The results for \xi bring up again the question of the soil properties at the soil-pile interface vs. those in the soil body.  We will discuss this in a later post.

References

Other than those in the original study, the following work was cited:

  • Santamarina, J.C., and Fratta, D. (1998) Introduction to Discrete Signals and Inverse Problems in Civil Engineering.  ASCE Press, Reston, VA.