2 views

Uploaded by Ibnu Riswan

With the rapid development of industrialization and
urbanization, people drive up demands for mineral
resources, and the consumption of mineral resources is
exploding. Nowadays, the digital prospecting has became
one of the major trends in metallogenic prognosis with the
development of information technology, more extensive
application of quantitative method and information
technology is demanded. The formation of mineral deposit
relates to

- tolerancing.pdf
- Helicopter Design Roadmap
- Finite Element Program in Autocad Vba
- 07 CMOST Tutorial
- Design 1
- Assignment 1
- Libro planificacion en recursos hidricoscompleto.pdf
- Intro
- Fea-Asme Secviii Div2 Seminar_190327
- FE-Einfuehrung
- Chapter 1 – One Variable Optimization
- staadpro-100916081053-phpapp01
- Finite Elements
- 2014 03 17 Jou Pub Front Stru Civ Eng Kim Et Al, Finite Element Modeling
- ME 1401 - FINITE ELEMENT ANALYSIS Two Marks Questions With Answers 2014 _ Vidyarthiplus (V+) Blog - A Blog for Students
- IFEM Solution Ch15
- Lab1
- reagan_uq
- Chapter_5.pdf
- 153-155NCICE-078

You are on page 1of 9

DATA UNCERTAINTY

By Andrew J. Graettinger,1 Associate Member, ASCE, and

Charles H. Dowding,2 Member, ASCE

ABSTRACT: Quantitatively directed exploration (QDE) employs a first-order Taylor series expansion to com-

bine sensitivity of a 3D finite-element model (FEM) and uncertainty in geologic data to calculate the variance

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

in project performance, which is employed to direct exploration. This approach is made practical by calculating

model sensitivity with direct differentiation of the engineering analysis code, thus producing sensitivity with a

single model run rather than multiple runs required by parameter perturbation. Uncertainty in subsurface data

is computed through two different extrapolation methods for comparison: kriging and conditional probability

(Bayesian updating). Although either of these methods can be employed in QDE, conditional probability is

required to quantifiably terminate exploration. The QDE framework is applicable to any subsurface analysis that

employs a 3D FEM. A case study illustrates the QDE approach, where settlement is the performance criterion,

and layer interface elevations are the uncertain geologic data. Additional boring locations identified by QDE

were placed where a combination of model sensitivity and subsurface uncertainty was the greatest, thus directing

exploration toward the building footprint and away from existing sampled points.

to provide interface elevation information at unsampled points.

Optimizing the location of the small number of borings em- Sensitivity is calculated efficiently through QDE by direct

ployed during site characterization is becoming increasingly derivative coding (DDC) (Wilson and Metcalfe 1985). In the

important because the need for more precise characteriza- example presented, sensitivity of calculated settlement is a

tion conflicts with the limited availability of exploration re- function of the sensitivity of the mesh to changes in interface

sources. Typically less than one-millionth of a site volume is locations multiplied by the sensitivity of the settlement cal-

sampled during exploration (Poeter and McKenna 1995). culated by 3D finite-element model (FEM) to changes in the

Therefore, engineering and geologic judgment are relied upon mesh. The derivatives are calculated at every node location

to leverage this information to direct exploration and prepare throughout the 3D model space. Because DDC allows for all

models for analysis. The need for a more rigorous, less sub- sensitivities to be calculated with a single model run, this

jective, probabilistic approach to site exploration for geotech- method is more computationally efficient than traditional sen-

nical engineering projects has been discussed for decades. This sitivity analyses produced by parameter perturbation or Monte

interest can be seen in articles dating back some 35 years, Carlo simulations. The QDE analysis of the case study pre-

including Casagrande (1965), Grivas (1977), Dowding (1978), sented in this paper required 120MB of memory and produced

National Research Council (1995), and Shackelford et al. approximately 200MB of output data; therefore, this approach

(1996). can be implemented on personal computers or workstations.

An efficient quantitative method for selecting boring loca- Because the QDE approach is based on a 3D FEM em-

tions called quantitatively directed exploration (QDE) is de- ployed to calculate project performance, it can be employed

scribed in this paper. This method combines (1) sensitivity of for a large variety of subsurface performance analyses. For

a 3D finite-element performance model; and (2) uncertainty of instance, EPA-sponsored research (Reeves et al. 1998) is cur-

input data, to locate the next boring at the position of greatest rently under way to apply this approach to ground-water flow

importancy. Importancy, which is the variance in calculated problems. QDE can be extended to other subsurface analyses

performance (Tomasko et al. 1987), is the product of the involving 3D FEM solutions, such as contaminant transport or

model sensitivity and input data uncertainty. Through the im- soil/rock-structure interaction. Project performance for each of

portancy matrix, QDE combines engineering analysis and ge- these cases is contingent on changes in the geologic interface

ologic characterization into a single, three-dimensionally dis- location and layer properties, which are dealt with indepen-

tributed variance parameter for directing exploration. dently in the QDE approach and in this paper.

This paper concentrates on interface uncertainty in the ex- The QDE method employs a first-order Taylor series ex-

ploration process and thus treats properties within layers as if pansion to produce an importancy matrix or variance in project

they are perfectly known, even though their uncertainty could performance (Ditmars et al. 1988). The importancy or variance

be incorporated within QDE. Effects of uncertainty within a in calculated performance is needed because neither the sen-

layer were not included as this subject has been discussed by sitivity (greatest near the largest loads) nor the uncertainty

others (Wilson and Metcalfe 1985; Baecher and Ingra 1986; (greatest at a distance from boring locations) alone indicates

Ditmars et al. 1988). Interface data from boring locations can the optimal location for the next boring. In the QDE method

be extrapolated by either kriging (Davis 1986; Isaaks and Sri- described herein, information gathered from the optimal bor-

ing location will yield the greatest reduction in calculated per-

1

Asst. Prof., Dept. of Civ. and Envir. Engrg., Univ. of Alabama, Tus- formance variance. The fundamental mathematics of QDE,

caloosa, AL 35487. based on a first-order Taylor series expansion, are described

2

Prof., Dept. of Civ. Engrg., Northwestern Univ., Evanston, IL 60208. later.

Note. Discussion open until April 1, 2000. To extend the closing date Application of QDE is described through an example to

one month, a written request must be filed with the ASCE Manager of demonstrate the method by which the next most important

Journals. The manuscript for this paper was submitted for review and

boring location is selected. To provide the proper setting for

possible publication on April 9, 1998. This paper is part of the Journal

of Geotechnical and Geoenvironmental Engineering, Vol. 125, No. 11, the case study, the QDE computational flowchart is described

November, 1999. 䉷ASCE, ISSN 1090-0241/99/0011-0959–0967/$8.00 first. The data flow structure and case study lay the foundation

⫹ $.50 per page. Paper No. 18090. for subsequent discussion of the computation of uncertainty in

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / NOVEMBER 1999 / 959

characterization, performance model sensitivity, and variances iance in calculated performance (importancy matrix, column

in calculated performance. Mathematical details of the data 4); and (3) a predetermined performance goal. Comparison of

extrapolation methods employed are introduced in the body the performance goal and the probability distribution of cal-

and are discussed in Appendix I. culated performance allows for the calculation of project re-

liability. If the reliability index is sufficient, given the current

DATA FLOW STRUCTURE FOR QDE subsurface uncertainty, then exploration stops. If it is insuffi-

cient, more exploration is required. This paper concentrates on

The computational flowchart for QDE is shown in Fig. 1. the details of the exploration process where the reliability of

Data output from one module or program is employed as input project success is insufficient and more exploration is required.

for the next. Computational steps normally associated with site

characterization are shown at the top, and those associated CASE STUDY OF DIRECTED EXPLORATION

with modeling or analysis of project performance are shown

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

at the bottom. The computational process begins at the top left Operation of QDE to locate the next boring, to optimally

with interface elevation data from the borings. These limited reduce the variance in surface settlement, is demonstrated with

interface data are then extrapolated to generate interface sur- actual field data. The site is located on thick alluvial deposits

faces with associated uncertainty throughout the site. Extrap- consisting of silty clay and dense sands overlying stiff to very

olated data are inherently uncertain because they are produced stiff clays (Finno 1996). The plan view of the site in Fig. 2

from a relatively small number of borings or sampling points shows the location of initial borings and the footprint of the

(Kulhawy and Trautmann 1996). The uncertainty matrix proposed two-story structure. Five existing soil borings (SB-

shown in the second column of Fig. 1 stores the uncertainty 1–SB-5) drilled to a depth of 15.25 m define the subsurface.

of the extrapolated 3D interface surfaces. A 0.75-m-thick concrete mat is designed to support interior

Moving down column 1 from the top left, the extrapolated equipment, and a 1.5-m-wide perimeter strip footing is de-

3D interface surfaces serve as input for the meshing program, signed to support the building frame; there are no interior col-

which creates nodal coordinates and element connectivity for umns. Design loads are 23 kPa for the mat and 48 kPa for the

the FEM. In addition, the meshing program distributes soil perimeter strip footings. These loads were averaged across the

properties, which for this example have zero uncertainty, to building footprint and applied as nodal loads in the FEM.

each of the mesh elements. Continuing to move down column The performance goal for this example was to restrain total

1, the mesh then serves as input for the 3D FEM settlement surface settlement of the structure. Other performance criteria

program that computes project performance (in this case nodal can be considered with the QDE approach. Both total and

displacements) in 3D space. differential settlement performance criteria have been evalu-

QDE expands upon the typical 3D settlement analysis with ated (Graettinger 1998). It was shown that these two criteria

the addition of a sensitivity analysis of the performance model each direct exploration to the same location because differ-

and meshing program by DDC, as shown by the dashed boxes ential settlement is a function of total settlement. The power

in Fig. 1. Following this new step, the derivative matrices are of the QDE approach is that both criteria as well as others can

multiplied together to produce the overall sensitivity matrix, be investigated rapidly.

shown in column 2. This sensitivity matrix is then multiplied Interface elevations of four geologic layers are shown by

by the uncertainty matrix using the Taylor series calculation, the cross section in Fig. 3(a). The dashed line in Fig. 2, labeled

shown in column 3, to produce the importancy matrix or var- A-A⬘, locates this representative cross section with respect to

iance in calculated performance. the building. Soil properties of Young’s modulus E and Pois-

The reliability index, shown in column 4, is employed to son’s ratio for each layer were determined from soil samples

determine when exploration is adequate. It is produced from and used as input for the 3D FEM. Table 1 shows soil param-

a combination of (1) the expected value of calculated perfor- eters for the four layers at this site. Although it is possible to

mance (predicted settlement, bottom of column 1); (2) the var- incorporate uncertainty of layer properties by adding another

FIG. 1. Flowchart of Subprograms (Boxed and Capitalized) and Data Matrices (Unboxed and Lower Cased) Required to Produce Im-

portancy Matrix Employed in Directed Exploration [after Graettinger and Dowding (1997)]

TABLE 1. Soil Layer Properties

Young’s modulusa E

Layer (kPa) Poisson’s ratio

(1) (2) (3)

Silty clay 6,420 0.3

Dense sand 26,350 0.3

Stiff clay 18,400 0.3

Very stiff clay 22,810 0.3

a

From constrained modulus and assumed Poisson’s ratio.

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

the focus of this paper that interface location is any more im-

portant than soil properties.

A spatial correlation structure based upon interface eleva-

tions at the case study borings was employed to extrapolate

interfaces. The correlation structure is determined by fitting

FIG. 2. Plan View of Case Study, which Includes Building Out- typical variogram shapes (Isaaks and Srivastava 1989) to the

line, Existing and Proposed Soil Boring Locations, and Location

site specific data of interface elevation differences and bore-

of Cross Section A-A⬘

hole separation distance. Different interfaces were considered

to be statistically independent. Both linear and Gaussian var-

iogram shapes were employed in this example. The Gaussian

shape is shown in the inset in Fig. 3(a). Vertical bars on the

variogram inset indicate the horizontal dimensions of the finite

elements. This spatial correlation is similar to those reported

in the literature for soils (Lacasse and Nadim 1996).

Total importancy, which is the variance of calculated surface

settlement, is employed to determine the location for the next

boring. As seen in Fig. 4, total importancy is a minimum at

sampled locations and at distances away from the proposed

building. Importancy is produced by first multiplying the sen-

sitivity of surface settlement to changes in interface elevation

by the associated elevation uncertainty and then summing

these values for all four interfaces to produce a total impor-

tancy or settlement variance at each x, y location.

Two approaches for directing exploration are available to

analyze the spatial distribution of total importancy. First, an

exploration philosophy that focuses on the ‘‘worst-case’’ as-

pect of the site directs exploration to the x, y location of largest

total site importancy. This is the location with the greatest

variance in calculated settlement. Second, a philosophy of

gaining ‘‘overall knowledge’’ at a site directs exploration by

minimizing the volume under the spatial distribution of total

importancy shown in Fig. 4.

For this example, the worst-case philosophy was employed

to identify the optimal location for the next soil boring, labeled

SB-6 in Fig. 5(a). Fig. 5(a) is a contour plot of the total im-

portancy surface shown in Fig. 4. The asymmetry of impor-

tancy results from the asymmetry of the building footprint

(Fig. 2), the interface elevation and variance (Fig. 3). Assumed

interface elevation data were added to the original site infor-

mation as if SB-6 were drilled, and the QDE analysis was

FIG. 3. (a) Cross Section A-A⬘ Including Building Location,

Geologic Layer Interfaces, and Interface Uncertainty Error Bars. rerun to produce a new total site importancy, shown in Fig.

Inset Shows Variogram Indicating How Interface Uncertainty Is 5(b). The same worst-case exploration philosophy used to lo-

Reduced near Boring Locations. (b) Cross Section A-A⬘ Show- cate SB-6 was then employed to locate the next sampling

ing Location of New Soil Boring SB-6 at 42.5 m, and Its Affect on point, SB-7, which is shown in Fig. 5(b).

Interface Elevation Uncertainty The significance of including ‘‘professional judgment’’ in

QDE without basing the next boring location solely on pro-

set of terms to the Taylor series expansion (Baecher and Ingra fessional judgment cannot be overemphasized. Subjectivity is

1981), layer properties are held constant in this example to only introduced in the initial or prior model. This prior model

focus attention on the spatial uncertainty of the interface lo- involves initial elevations, the correlation structure, and initial

cations. Consideration of a standard 1D settlement equation variances. From then on, determination of the next boring lo-

indicates that the contribution to the variance of settlement due cation is calculated. Furthermore, the intuitively obvious lo-

to soil properties Cc is the same as it is due to the layer height cation of new borings SB-6 and SB-7 within the building foot-

H. Because soil properties and layer thickness have a similar print is a demonstration of the validity of the QDE approach.

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / NOVEMBER 1999 / 961

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

FIG. 4. Relationship between Borings (SB-1–SB-5) and the Square Root of the 3D Total Importancy Distribution Employed to Direct

Exploration

for exploration, none were employed in this calculation.

TERMINATION OF EXPLORATION

As shown in Fig. 1, the QDE method can be expanded to

include a quantifiable conclusion to exploration through the

reliability index. Termination decisions are made by compar-

ing the probability distribution of calculated performance to a

predetermined performance goal, from which a reliability in-

dex or probability of project success is calculated. This pro-

cess, called reliability-based exploration (RBE), has been suc-

cessfully applied to two synthetic case studies (Graettinger

1998). The two cases differ only in the geometry of the sub-

surface layers. One site involved an abruptly pinching soft clay

layer directly below the proposed structure, whereas the sec-

ond site’s clay layer gradually tapers across the entire site.

Because the case studies are synthetic, the QDE modeled in-

terfaces could be compared with the predetermined interfaces.

During the evaluation of exploration sufficiency at the two

synthetic sites the following results were identified. First, to

accurately model the abruptly pinching geometry, a directional

variogram was required. Second, the site with an abruptly

pinching geometry required more borings to define project per-

formance with the same confidence as for the site with a grad-

ually tapering geometry. The study also showed that regardless

of the number of additional borings at the pinching site, the

desired reliability could not be met. In this situation the design

must be changed or the objectives relaxed.

AND UNCERTAINTY

Two methods of data extrapolation, kriging and conditional

probability or Bayesian updating, were tested and shown to

direct exploration to the same location. Despite the similarity

of the results in this case, each approach represents a different

philosophical view of extrapolation (Freeze et al. 1990). Krig-

ing is based on classical statistics and does not allow for sub-

jective prior estimates of interface elevation, whereas condi-

tional probability does. Both methods were employed to

extrapolate data from the borings to a regularized finite-ele-

ment grid.

FIG. 5. (a) Contour Map of the Square Root of Total Impor- Uncertainty in Extrapolated Data

tancy (in mm) Indicating Location of Next Boring SB-6; (b) New

Importancy Distribution after Inclusion of Data from SB-6, along Kriging and conditional probability calculate interface ele-

with Location of Next Soil Boring SB-7 vations and the associated spatially distributed uncertainty.

962 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / NOVEMBER 1999

Each interface is an extrapolated 3D surface; therefore, the variance produced by kriging is not the true variance but rather

associated uncertainty is also 3D and is assumed normally dis- the error variance. Although the assumption that the error var-

tributed about the calculated elevation. Fig. 6 shows the un- iance is the true variance allows kriging to be employed for

certainty in the stiff clay-very stiff clay interface, which de- QDE, it should not be employed for ending exploration

creases at the five sampled locations. These uncertainties are through RBE. Because conditional probability can incorporate

seen along cross section A-As⬘ in Figs. 3(a and b) as a series the covariance of input data it will produce a more accurate

of vertical bars. Longer error bars indicate a greater uncer- variance and, therefore, is the preferred extrapolation method

tainty in the interface location. Uncertainty along cross section for the QDE and RBE approaches.

A-A⬘ in Fig. 3(a) decreases in three regions that are near bor-

ings SB-3–SB-5. As is expected, uncertainty increases with Kriging

distance from these borings as shown by the variogram inset

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

in Fig. 3(a) (Isaaks and Srivastava 1989). The ground surface Kriging requires the assumption that the calculated error

interface has low uncertainty because it is assumed to be a variance is the true variance of unsampled interface elevations

horizontal domain boundary in this example. to produce a probabilistic subsurface model (Smith 1989). In

As seen in Fig. 3(b), the uncertainty distribution is greatly kriging, the covariance function is employed to determine an

affected by the addition of the new soil boring SB-6 at the optimal set of weights that are used in the calculation of the

location of greatest importancy. In this example, SB-6 inter- best estimate of interface elevation and the associated uncer-

sects each interface, thus reducing uncertainty to zero at the tainty (Davis 1986; Isaaks and Srivastava 1989). The set of

sampled location. The uncertainty near the new boring is re- weights for calculating the best estimate of elevation at a spe-

duced for all interfaces, but is most easily observed on the stiff cific x, y location is the product of the inverse of a matrix

clay-very stiff clay interface in Fig. 3(b). The location of this containing the covariance between sampled elevations multi-

interface was relatively uncertain before SB-6 data were added plied by a vector containing the covariance between the sam-

and is still uncertain except for the area near the new soil pled elevations and the elevation being estimated. The weight-

boring. ing vector takes into account two important aspects of the

estimation process: (1) spatial clustering of the known (sam-

Comparison of Kriging and Conditional Probability pled) elevation data; and (2) spatial separation between the

known (sampled) and unknown (unsampled) elevation (Isaaks

The two extrapolation methods investigated in this study, and Srivastava 1989). The same set of weights employed to

kriging and conditional probability (Bayesian updating), em- calculate the best estimate of interface elevation is also em-

ploy covariance functions to calculate mean interface eleva- ployed to produce the error variance, which must be assumed

tions and uncertainty of interface elevations throughout the in this approach to be equal to the uncertainty in interface el-

model space. A covariance function in the form of a vario- evation. Details of the kriging approach are described in Ap-

gram, which defines the relationship between interface eleva- pendix I.

tions and the distance separating those elevations, is shown in

the inset of Fig. 3(a). Because the covariance function is in- Conditional Probability

dependent of the extrapolation method, the same covariance

function was employed with each method. Thus it is not sur- Extrapolation through conditional probability (Bayesian up-

prising that the two extrapolation methods incorporated in the dating) requires a prior or initial estimate of the probabilistic

QDE method each yield the same location of greatest total subsurface (Gelman et al. 1995). For this study, a prior inter-

importancy. face elevation was assumed to be a flat surface located at the

The extrapolated interface elevation data produced with the average value of the five sampled elevations, and the initial

conditional probability approach were within 2% of the values variance was assumed to be half the maximum variance (or

produced by kriging. The variance produced by conditional variogram sill) of the sampled locations. Unlike kriging, the

probability, which is the main diagonal of the covariance ma- conditional probability method incorporates initial engineering

trix, was within 13% of the values produced by kriging. The judgment or ‘‘soft data,’’ by allowing the interface elevations

FIG. 6. 3D Spatial Uncertainty Associated with Stiff Clay-Very Stiff Clay Interface, which Declines at Boring Locations

to be adjusted in the prior model to ‘‘better’’ represent the duced from both of the enhanced programs were multiplied

site. The same covariance function employed in kriging was together to produce the 3D sensitivities.

used to generate the prior covariance matrix, which describes Model sensitivity, or d(surface settlement)/d(interface ele-

the spatial relationships between every pair of points on a spe- vation), is a product of the derivatives from both the meshing

cific interface. and the 3D finite-element programs, as seen in Fig. 1, column

In the conditional probability approach, the prior model is 2. In this 3D case study there are 676 surface nodes and 3,380

updated with ‘‘hard’’ data from the sampled points through a interface elevation nodes; therefore, (676 ⫻ 3,380) = 2.28 ⫻

series of matrix manipulations similar to those employed in 106 sensitivities are required to describe every d(surface set-

kriging, as shown in Appendix I. Unlike conditional probabil- tlement)/d(interface elevation) derivative.

ity, the kriging approach does not produce a posterior covar- Sensitivity is calculated for every node in the finite element

iance matrix; therefore, an uncorrelated version of the Taylor mesh. A sense of the scale of the finite-element mesh can be

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

series approximation is employed to calculate importancy gained from the inset in Fig. 2 and the tick marks on the

when input data are extrapolated by kriging. The effect of the variogram in Fig. 3(a). The elements are approximately one-

covariance terms on the QDE approach is discussed in the fifth the correlation length. The height of each element varies

section on importancy. to represent the spatially varying interface elevations. The 3D

Conditional probability and Bayesian updating approaches sensitivity distributions for each interface at the site are shown

are currently employed to direct exploration; however, their in cross section A-A⬘ in Fig. 7(a). In this figure, sensitivity is

use is based upon site uncertainty only and do not incorporate the ratio of a change in z displacement of a surface node,

the sensitivity of project performance as does QDE and RBE. divided by an increase in the interface elevation. A negative

For instance, Johnson (1996) employed a Bayesian/geostatis- sensitivity corresponds to greater calculated surface settlement

tical adaptive sampling approach, which uses uncertainty in caused by an increase in interface elevation, whereas positive

soil contaminant levels to direct exploration. Johnson’s con- sensitivity corresponds to less surface settlement.

taminated soil example has no associated performance model, Sensitivity is relatively unaffected by the addition of new

only a spatial material property (contamination); therefore, in- interface elevation data in this example, as shown by the sim-

put data uncertainty alone is sufficient to direct exploration in ilarity between the original sensitivity [Fig. 7(a)] and the sen-

that case. sitivities after synthetic data were added from SB-6 as seen in

Fig. 7(b). Interface sensitivities increase or decrease only when

SENSITIVITY OF PERFORMANCE MODEL FROM DDC the best estimate of an interface location moves closer to or

away from the load. As seen in Figs. 3(a and b), the best

Sensitivity of the performance model, which for this ex- estimate of interface locations along cross section A-A⬘ are

ample is defined as a change in surface settlement divided by approximately the same; therefore, the sensitivities will remain

a change in each interface elevation, was produced by directly approximately the same. The most sensitive subsurface inter-

differentiating the existing programs. This process, DDC, em- face in the site is the silty clay-dense sand interface because

ploys the chain rule of calculus to produce accurate derivatives

with a single model run. Two computer programs, a meshing

and a 3D FEM program, were employed to calculate surface

settlement from interface elevations. Both programs were re-

written to produce the original output plus gradients that de-

scribe model sensitivity for any layered geometry. DDC is un-

dertaken only once for each program, but model sensitivities

are re-evaluated after each new boring is drilled because sen-

sitivities change if the best estimate of interface elevation

changes. Programming of the derivatives was accomplished

either by hand coding or by employing ADIFOR 2.0 (Bischof

et al. 1995), an automated derivative coding program. This

method of evaluating model sensitivities was compared to the

two-sided parameter perturbation method and was shown to

be 125 times faster (Graettinger 1998).

Horwedel et al. (1992) states that existing codes can be re-

programmed to produce the original output plus derivatives

thus eliminating the need to perturb input variables to evaluate

model sensitivities. By analyzing each line of code and apply-

ing the chain rule of calculus, an entire program can be sys-

tematically differentiated. Manual reprogramming of existing

code is a tedious process that is prone to errors and in many

cases takes much longer than the original coding (Bischof et

al. 1995); thus software was developed to automatically re-

program existing code. ADIFOR 2.0 (Automatic Differentia-

tion of FORTRAN) is such a preprocessor that automatically

rewrites operating FORTRAN 77 programs to generate deriv-

ative enhanced codes. ADIFOR reads FORTRAN code as in-

put and produces a new program as output that retains the

original code’s functionality while adding the ability to pro-

duce the derivative of dependent output variables with respect

to independent input variables. In this study ADIFOR 2.0 was

employed to enhance the meshing program and specific sub-

routines of the finite-element program (Johnston 1981; Finno FIG. 7. (a) and (b) Change in Sensitivity with Addition of New

1983). In addition, hand coding was used to differentiate cer- Boring SB-6; (c) and (d) More Significant Change in Importancy

tain aspects of the finite-element program. The derivatives pro- with Addition of New Boring SB-6

(1) it is closest to the load; and (2) compressibility of the silty comparing Figs. 3(a and b). Therefore, performing the com-

clay and dense sand layers are very different. Thus, a change putation described in (1) at SB-6 (42.5 m along cross section

in this interface elevation will greatly affect the height of the A-A⬘) results in a low importancy or variance in calculated

closest, most compressible layer. surface settlement. The largest reduction in interface impor-

tancy is with the silty clay-dense sand interface as seen by

IMPORTANCY GOVERNS PERFORMANCE comparing Figs. 7(c and d) at SB-6.

DRIVEN DECISIONS

EFFICIENCIES OF DIRECT DIFFERENTIATION AND

Neither uncertainty in interface location nor model sensitiv- TAYLOR SERIES EXPANSION

ity alone can adequately direct or evaluate exploration. In the

QDE approach these two considerations are combined through Combining DDC and first-order Taylor series expansion

dramatically reduces run times compared to employing param-

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

distribution, which mathematically is the variance in computed eter perturbation or Monte Carlo methods. The Monte Carlo

surface settlement. The importancy at any surface node is in- method requires numerous site realizations, each of which are

fluenced by all interface elevations at the site because each run through the engineering model. Alternatively the Taylor

interface elevation affects surface settlement in this 3D ex- series with DDC requires only one model run for any subsur-

ample. Unlike sensitivity, which changes only when the best face geometry. During this run the sensitivity, performance

estimate of the interface elevation changes after the addition variance, and the model performance are all calculated.

of a new boring, importancy responds significantly to the new The Taylor series with the DDC method has successfully

information because of the reduction of uncertainty around the been employed to calculate the sensitivity and probability dis-

boring. tributions for performance models with 10,000 dependent var-

Interface importancy distributions, shown in Figs. 7(c and iables produced from 4,000 correlated input variables in a mat-

d), are calculated for each interface through a first-order Taylor ter of hours. The accuracy and computational efficiency of the

series expansion, as shown by (1) (Harr 1996), which is for Taylor series method based on DDC has been compared to

correlated input data Monte Carlo simulations for the 3D subsurface investigation

冘冘冉 冊冉 冊

n n

described herein. Results for both correlated and uncorrelated

⭸fk ⭸fk Monte Carlo simulations showed accurate agreement to the

var[Yk] ⬇ Cov[x̄i, x̄j] (1)

i=1 j=1 ⭸x̄i ⭸x̄j results produced with a single model run of the Taylor series

expansion method. The Taylor series method required only 2.5

where var[Yk] = variance or interface importancy of computed h of computer time on a 160-MHz machine; however, 10,000

surface settlement at node k. For this case study there are 676 uncorrelated Monte Carlo simulations required approximately

surface nodes. (⭸fk /⭸x̄i) is the sensitivity of surface node k to 10 days on a 200-MHz machine, and 10,000 correlated Monte

the elevation of interface node i ; because there are 676 surface Carlo simulations required some 12 days with Choleski de-

nodes and 676 elevation nodes on each interface, there are composition (Ripley 1987).

over (676 ⫻ 676) = 4.50 ⫻ 105 surface sensitivities for each

interface. Cov[x̄i, x̄j] is the covariance between elevations of CONCLUSIONS

interface nodes i and j. There are 6762 pairs of nodes on each

interface or over 4.50 ⫻ 105 covariance terms. Because only The QDE approach presented herein employs a first-order

the variance or main diagonal of the covariance matrix is pro- Taylor series expansion to combine the sensitivity of a 3D

duced from kriging, there is no updated covariance matrix. FEM and uncertainty in extrapolated geologic characterization

Therefore, the uncorrelated version of the first-order Taylor to calculate the importancy or variance in project performance.

series expansion must be employed. Although the performance model for this example involved

The Taylor series expansion method for calculating perfor- computed surface settlement, the generality of the finite-ele-

mance variance has limitations. First, the derivatives are cal- ment computation allows use of the QDE approach in a variety

culated at the expected value of interface elevation; therefore, of performance measures ranging from soil/rock-structure de-

it there is a large uncertainty associated with the interface el- formation to ground-water flow. Though the illustrated ex-

evation the derivative may not be equal to the derivative of ample focused only on the effect of uncertainty in geologic

the true interface elevation (Harr 1996). Second, in this ex- interface elevations, the QDE framework can include the effect

ample the interface elevations on one interface were assumed of uncertainty in layer properties. Based upon these accom-

to be statistically independent of the other interfaces in the plishments and the study presented, the following conclusions

subsurface. Finally, the Taylor series expansion method re- are advanced:

quires knowledge of the uncertainty associated with each input

parameter, which in this example is produced during interface • Both sensitivity of the project performance model and un-

extrapolation. certainty in the geologic input data must be combined to

Fig. 7(c) shows importancy values var[Yk], for each interface quantitatively direct exploration through calculation of

along cross section A-A⬘ before the addition of SB-6 elevation importancy, or variance, in the project performance. This

data. The large importancy values for the silty clay-dense sand settlement example demonstrated a method of accounting

interface, shown in Fig. 7(c), indicate that this interface is the for uncertainty of 3D interface elevations. Additional bor-

most significant to explore because it is uncertain and it de- ing locations that were identified by QDE were intuitively

termines the thickness of the closest and most compressible obvious, and thus confirm the adequacy of the approach.

layer. These individual interface importancy values are • No subjective judgment beyond acceptance of the initial

summed vertically at a specific location (x, y) to produce the geologic model and correlation structure is required. Ex-

total importancy value. The distribution of total importancy trapolation of interface elevations from known values at

for the site is seen in Figs. 4 and 5. borings through either kriging or conditional probability

The interface importancies are greatly influenced by the ad- allows for calculation of interface elevation and uncer-

dition of soil boring SB-6, as shown in Fig. 7(d). Although tainty at unsampled locations. However, conditional prob-

the interface sensitivities in Figs. 7(a and b) remained rela- ability provides the full covariance matrix to completely

tively constant in this example, the input data uncertainty at describe the probabilistic subsurface, which is necessary

the new boring location is significantly reduced as seen by for the correlated version of the first-order Taylor series

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / NOVEMBER 1999 / 965

expansion as well as termination of exploration through couraged to investigate introductory kriging literature for fur-

calculation of the reliability of project performance. ther details on this process (Davis 1986; Isaaks and Srivastava

• Direct differentiation of the finite-element code allows 1989).

sensitivities to be computed with a single pass through Kriging also produces an estimate of error variance, which

the program, which significantly reduces the computa- is assumed, as by many others, to represent the true spatial

tional effort compared with either parameter perturbation variance in the parameter being estimated. The error variance

or Monte Carol simulation. Because direct differentiation in estimated interface elevation var[x̄], produced with the krig-

affects only the meshing and finite-element programs and ing process, is shown as follows:

冘

not the data, it is neither site nor structure dependent and m

only needs to be undertaken once for the code to be em-

var[x̄] = var[X] ⫺ wi Cov[Xi, x] ⫹ (7)

ployed for any number of sites and facilities. Efficient i=1

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

calculation of performance variance or importancy result- where, for this paper, var[X] = maximum variance in interface

ing from the direct differentiation of the analytical code elevation set by the magnitude of the covariance function; wi

allows QDE to be employed in adaptive real-time explo- = weight factor; Cov[Xi, x] = spatial covariance between the

ration. known elevation locations and the elevation being estimated;

and = Lagrange parameter. All the terms employed in the

APPENDIX I. INPUT DATA EXTRAPOLATION kriging process to estimate elevation and variance are either

MATHEMATICAL BACKGROUND actual data from the site or are produced from the covariance

The QDE approach presented in this case study examined function or variogram model employed to model the spatial

two extrapolation methods, kriging and conditional probability correlation of the data.

(Bayesian updating), to generate a regularized grid of interface The conditional probability approach employed in this ex-

elevation estimates and associated uncertainty. For this ex- ample is for multivariate, normally distributed data. Utilizing

ample, both extrapolation methods directed exploration to the only the limited information retrieved from sampling points,

same x, y location and produced the same value of settlement three sets of data are needed to begin this approach. First, the

variance when only the diagonal terms of the posterior covar- actual interface elevations from the five boring locations are

iance matrix from conditional probability were employed in stored in a vector V. Second, a vector E(U ) is produced that

the Taylor series calculation. This appendix provides a brief contains a prior estimate of mean elevation at each interface

look at these two methods and examines their similarities. grid point. For this example the interface was initially assumed

Kriging extrapolates interface elevation from known loca- flat at the average elevation of the five sampled points. Finally,

tions at borings to unsampled locations through a weighting an informed prior covariance matrix Cov(U ), is calculated by

process described below. The kriging procedure estimates in- employing a function that represents the covariance relation-

terface elevation x̄ at an unsampled location by solving (2) ships between the known interface elevations at boring loca-

(Isaaks and Srivastava 1989) tions. From these matrices, entirely based on site information,

the prior estimates E(U) and Cov(U) are updated to posteriors

冘

m

E(U兩V) and Cov(U兩V) through the following equations (Gel-

x̄ = wi Xi (2) man 1995):

i=1

where Xi = known interface elevation at boring i; and wi = E(U兩V ) = E(U ) ⫹ Cov(V, U )Cov(V )⫺1(V ⫺ E(V )) (8)

weighting factor based on the covariance between known in- Cov(U兩V ) = Cov(U ) ⫺ Cov(V, U )Cov(V )⫺1Cov(U, V ) (9)

terface elevations and the interface elevation being estimated

at an unsampled location. The weighting factor is calculated E(U兩V) = updated vector of interface elevation given the five

by known elevations at the borings. Cov(V, U) is a subset of the

full covariance matrix describing the interface and is the co-

w = C ⫺1D (3)

variance between the known elevations and the remaining in-

再冎

where terface elevations being estimated. In this example the Cov(V,

U) matrix is 676 ⫻ 5. Cov(V)⫺1 is again a subset of the full

w1 covariance matrix and is the inverse of the covariance between

⭈⭈

w= ⭈ (4) the known interface elevations. Cov(V)⫺1 is a 5 ⫻ 5 matrix

wm for this example. E(V) is a subset of E(U) and holds the prior

冋 册

elevation estimates at the five boring locations. In (9),

Cov(U兩V) is the updated covariance matrix, which has dimen-

Cov[X1, X1] ⭈⭈⭈ Cov[X1, Xm] 1

⭈⭈ ⭈⭈ ⭈⭈ sions of 676 ⫻ 676 terms. Cov(U) is the prior covariance

⭈ ⭈ ⭈ ⭈⭈⭈ matrix, 676 ⫻ 676, calculated by the covariance function, and

C= (5)

Cov[Xm, X1] ⭈⭈⭈ Cov[Xm, Xm] 1 Cov(U, V) is the transverse of Cov(V, U).

再 冎

1 ⭈⭈⭈ 1 0 Comparing these two extrapolation methods for the estimate

Cov[X1, x̄] of interface elevation reveals many similarities. The same var-

⭈⭈ iogram or covariance function is employed in both methods

D= ⭈ (6) for each interface. Therefore, the covariance terms between

Cov[Xm, x̄]

1 pairs of points in each method will be the same. The C matrix

in kriging [(5)], ignoring the Lagrange parameter, is the

The weight vector seen in (4) is a product of the inverse of a Cov(V) matrix in conditional probability. The D vector in krig-

matrix containing the covariance Cov[Xm, Xm] between known ing [(6)] is one row of the Cov(V, U) matrix in conditional

interface elevations at boring locations [(5)] multiplied by a probability, and as successive grid points in kriging are esti-

vector of covariance Cov[Xm, x̄] between known interface el- mated successive rows of the Cov(V, U) matrix match the D

evations at borings Xi and the interface elevation being esti- vector. These two matrices are inversely multiplied together in

mated at an unsampled location x̄ [(6)]. Both the C and D each method to produce a weighting factor. Kriging multiplies

matrices have additional terms for the Lagrange parameter the weighting factor by the known elevations at sampled lo-

that is needed to solve the kriging equation. Readers are en- cations. Conditional probability multiplies the weighting vec-

966 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / NOVEMBER 1999

tor by a vector of the difference between prior elevation esti- Graettinger, A. J., and Dowding, C. H. (1997). ‘‘Quantitative site explo-

mates and the known elevations at the boring locations and ration directed by interface location uncertainty.’’ Proc., 5th Great

Lakes Geotech./Geoenvir. Conf., Site Characterization for Geotech.

then adds this to the prior elevation estimate. Due to these and Geoenvir. Problems, University of Michigan, College of Engrg.,

similarities, for this example, interface elevations were ap- Ann Arbor, Mich., 117–131.

proximately equal. Grivas, D. A. (1977). ‘‘Probability theory and reliability analysis in geo-

technical engineering.’’ Rep. of an NSF-Sponsored Workshop at Rens-

ACKNOWLEDGMENTS selaer Polytechnic Institute, Rensselaer Polytechnic Inst., New York.

Harr, M. E. (1996). Reliability-based design in civil engineering. Dover,

The writers wish to thank Prof. G. Baecher of the Department of Civil New York.

Engineering, University of Maryland, College Park, Md., Drs. J. Ditmars, Horwedel, J. E., Raridon, R. J., and Wright, R. Q. (1992). ‘‘Automated

G. Williams, and R. Johnson of Argonne National Laboratories, Argonne, sensitivity analysis of an atmospheric dispersion model.’’ Atmospheric

Ill., and Profs. H. Reeves and T. Igusa at Northwestern University, Ev- Envir., 26A(9), 1643-1649.

Downloaded from ascelibrary.org by UNIVERSITI TEKNOLOGI PETRONAS on 03/04/15. Copyright ASCE. For personal use only; all rights reserved.

anston, Ill. Their advice has been extraordinarily helpful, and the writers Isaaks, E. H., and Srivastava, R. M. (1989). An introduction to applied

deeply appreciate their assistance and support. geostatistics. Oxford University Press, New York, 278–322.

Johnson, R. L. (1996). ‘‘A Bayesian geostatistical approach to the design

of adaptive sampling programs.’’ Geostatistics for environmental and

APPENDIX II. REFERENCES geotechnical applications, ASTM STP 1283, ASTM, West Consho-

hocken, Pa.

Baecher, G. B., and Ingra, T. S. (1981). ‘‘Stochastic FEM in settlement Johnston, P. R. (1981). ‘‘Finite element consolidation analysis of tunnel

predictions.’’ J. Geotech. Engrg. Div., ASCE, 107(4), 449–463. behavior in clay,’’ PhD thesis, Stanford Univ., Stanford, Calif.

Bischof, C., Carle, A., Khademi, P., Maure, A., and Hovland, P. (1995). Kulhawy, F. H., and Trautmann, C. H. (1996). ‘‘Estimation of in-situ

‘‘ADIFOR 2.0 user’s guide (revision C).’’ Tech. Memo. No. 192 and uncertainty.’’ Uncertainty in the geologic environment: From theory to

Ctr. for Res. on Parallel Computation Tech. Rep. CRPC-95516-S, practice, Geotech. Spec. Publ. No. 58, ASCE, New York, 269–286.

Mathematics and Comp. Sci. Div., Argonne National Laboratory, Ar- Lacasse, S., and Nadim, F. (1996). ‘‘Uncertainties in characterizing soil

gonne, Ill. properties.’’ Uncertainty in the geologic environment: From theory to

Casagrande, A. (1965). ‘‘Role of the ‘calculated risk’ in earthwork and practice, Geotech. Spec. Publ. No. 58, ASCE, New York, 49–75.

foundation engineering.’’ J. Soil Mechanics and Foundations Div., National Research Council. (1995). Probabilistic methods in geotechnical

ASCE, 91(SM4), 1–40. engineering. Board on Energy and Environmental Systems, National

Davis, C. D. (1986). Statistics and data analysis in geology. Wiley, New Research Council, Washington, D.C.

York. Poeter, E. P., and McKenna, S. A. (1995). ‘‘Reducing uncertainty asso-

Ditmars, J. D., Baecher, G. B., Edgar, D. E., and Dowding, C. H. (1988). ciated with ground-water flow and transport prediction.’’ Ground Wa-

‘‘Radioactive waste isolation in salt: A method for evaluating the ef- ter, 33(6), 899–904.

fectiveness of site characterization measurements.’’ U.S. Department Reeves, H. W., Dowding, C., and Igusa, T. (1998). ‘‘An efficient relia-

of Energy, Civilian Radioactive Waste Management Salt Repository bility based approach to aquifer remediation design’’ EPA Res. Proj.,

Project Office, available from National Technical Information Service, 98 NCERQA, EPA, Washington, D.C.

U.S. Department of Commerce, Springfield, Va. Ripley, B. D. (1987). Stochastic simulation. Wiley, New York, 237.

Dowding, C. H., ed. (1978). Site characterization and exploration. Proc., Shackelford, C. D., Nelson, P. P., and Roth, M. J. S., eds. (1996). Un-

Geotech. Div., ASCE, New York. certainty in the geologic environment: From theory to practice, Geo-

Finno, R. J. (1983). ‘‘Response of cohesive soil to advanced shield tun- tech. Spec. Publ. No. 58, ASCE, New York.

neling,’’ PhD thesis, Stanford Univ., Stanford, Calif. Smith, A. D. (1989). ‘‘Computerized modeling of geotechnical strati-

Finno, R. J. (1996). ‘‘LIGO report.’’ Internal Northwestern University graphic data,’’ PhD thesis, Massachusetts Institute of Technology,

document. Civ. Engrg. Dept., Northwestern University, Evanston, Ill. Mass.

Freeze, R. A., Massmann, J., Smith, L., Sperling, T., and James, B. Tomasko, D., Reeves, M., Kelley, V. A., and Pickens, J. F. (1987). ‘‘Pa-

(1990). ‘‘Hydrogeological decision analysis: 1. A framework.’’ Ground rameter Sensitivity and Importance for Radionuclide Transport in Dou-

Water, 28(5), 738–766. ble-Porosity Systems.’’ Proc., Conf. on Geostatistical, Sensitivity, and

Gelman, A., Carlin, J. B., Stern, H. S., and Rubin, D. B. (1995). Bayesian Uncertainty Methods for Ground-Water Flow and Radionuclide Trans-

data analysis. Chapman & Hall, London, 478–479. port Modeling, Battelle Press, Columbus, Ohio, 297–321.

Graettinger, A. J. (1998). ‘‘Reliability-based exploration: A quantitative Wilson, J. L., and Metcalfe, D. E. (1985). ‘‘Illustration and verification

method for evaluating and directing subsurface characterization,’’ PhD of adjoint sensitivity theory for steady state groundwater flow.’’ Water

thesis, Northwestern University, Evanston, Ill. Resour. Res., 21(11), 1602–1610.

- tolerancing.pdfUploaded by86sathesh
- Helicopter Design RoadmapUploaded byKaradias
- Finite Element Program in Autocad VbaUploaded byAmbrish Singh
- 07 CMOST TutorialUploaded byAnonymous Wo8uG3ay
- Design 1Uploaded byMuhammad Faisal Mahmod
- Assignment 1Uploaded byAnung Grahito
- Libro planificacion en recursos hidricoscompleto.pdfUploaded bySamuelson C
- IntroUploaded byआकाश संगीता प्रकाश पवार
- Fea-Asme Secviii Div2 Seminar_190327Uploaded byshm8324440
- FE-EinfuehrungUploaded byhdlchip
- Chapter 1 – One Variable OptimizationUploaded bykim haksong
- staadpro-100916081053-phpapp01Uploaded byHalomoan Harahap
- Finite ElementsUploaded bysonik2099
- 2014 03 17 Jou Pub Front Stru Civ Eng Kim Et Al, Finite Element ModelingUploaded bykristinakristina1234
- ME 1401 - FINITE ELEMENT ANALYSIS Two Marks Questions With Answers 2014 _ Vidyarthiplus (V+) Blog - A Blog for StudentsUploaded byibrahim
- IFEM Solution Ch15Uploaded bySajjad Ahmad
- Lab1Uploaded bywerwerwer
- reagan_uqUploaded byLeni Marlina
- Chapter_5.pdfUploaded byTATATAHER
- 153-155NCICE-078Uploaded byDewa Gede Soja Prabawa
- 00017037Uploaded byvbirolini
- Team Resumes Visualcv ResumeUploaded byasdfg ali
- Topic 1-Intro to Risk Management-English 2009Uploaded byDhea Yulianti P
- me471s03_lec07Uploaded byHasen Bebba
- Merton CaseUploaded bySomdeb Banerjee
- If EmUploaded byapi-3700440
- Lab 1Uploaded byCristian Camilo Sanchez
- schpdfgddgdgggUploaded bybhagathnagar
- fosters_aUploaded byRajarshi Chatterjee
- KMSCFDWeb Assignment Module5Uploaded bysadiksnm

- Chap3.pptxUploaded byIbnu Riswan
- hokmabadi2015Uploaded byIbnu Riswan
- Modelling of concrete using abaqus- nonlinear analysisUploaded byPrashanth Shyamala
- hokmabadi2015.pdfUploaded byIbnu Riswan
- ConcreteUploaded byalditikos
- 1. Ahmed 2012-Probabilistic Analysis of Strip Footings Resting on a Spatially Random Soil Using Subset Simulation ApproachUploaded byIbnu Riswan
- scnotes (1)Uploaded by2420570
- 2. Estimating Wall Deﬂections in Deep ExcavationsUploaded byIbnu Riswan
- 1-s2.0-S0266352X1400233X-mainUploaded byIbnu Riswan
- A case study for assessing uncertainty in local-scale regulatory air quality modeling applications.pdfUploaded byIbnu Riswan
- Quantitatively Directed Sampling for Main Channeland Hyporheic Zone Water-quality ModelingUploaded byIbnu Riswan
- Uncertainty Analysis in Air Dispersion ModelingUploaded byIbnu Riswan
- Reliability-based Design and Its Complementary Role to Eurocode 7 Design ApproachUploaded byIbnu Riswan
- StochBioChapter3Uploaded byVolodja
- Random Filed vs Monte CarloUploaded byIbnu Riswan
- Reliability-based design for transmission line structure foundations.pdfUploaded byIbnu Riswan
- Simulation of Non-Gaussian Processes Using Fractile CorrelationUploaded byIbnu Riswan
- Learning Bayesian network structure Towards the essential graph by integer linear programming tools.pdfUploaded byIbnu Riswan
- Modelling the Effects of Vegetation on Stability of Slopes.pdfUploaded byIbnu Riswan
- Stochastic Analysis of Structures in Fire by Monte Carlo SimulationUploaded byIbnu Riswan
- Using Abaqus Reliability Analysis Directional Sim 2008 FUploaded byIbnu Riswan
- 1. 2003-Bearing-capacity Prediction of Spatially Random c φ SoilsUploaded byIbnu Riswan
- 1. 2015-Effect of 2-D Random Field Discretization on Failure Probability and Failure Mechanism in Probabilistic Slope Stability [Model for Sentence]Uploaded byIbnu Riswan
- 1. 2011-Probabilistic Analysis and Design of a Strip Footing.pdfUploaded byIbnu Riswan
- Tutorial 6 - Copy.pdfUploaded byflorinelvv
- Error Evaluation of Three Random Field GeneratorsUploaded byIbnu Riswan
- Appc Soil Properties 718Uploaded bypinkuru
- Comparison of Three Covariance Functions on Stochastic Settlement Analysis of Shallow Foundation in Anisotropy FieldUploaded byIbnu Riswan

- Encyclopedia+of+Medical+Decision+MakingUploaded byAgam Reddy M
- Slide TutorialManual2Uploaded byrlpr
- 2013 Case Study AnsUploaded bypranav4560
- Qmsys Gum EnUploaded byAYKUT
- Lecture1_Introduction_ RobustControlTheory.pdfUploaded byajayaghimire
- Belt Weightometer VOAUploaded byquinteroudina
- DOE Wizard - Variance Component DesignsUploaded byAnonymous FZNn6rB
- GEOVIA WHITTLE THE WORLD´S MOST TRUSTED STRATEGIC MINE PLANING SOFTWAREUploaded byJOSE CARLOS MAMANI VALERIANO
- Quantifying in situ and modeling net nitrogen mineralization from soil organic matter in arable cropping systems - Hugues Clivot, , Bruno Mary, Matthieu Valé, Jean-Pierre Cohan, Luc Champolivier, François Piraux, François Laurent, Eric JustesUploaded byINRA - AGIR
- Kurt-ElliUploaded byravikk_ge
- Sunco QuestionUploaded byVarun Hknz
- Examples For abaqusUploaded byAswin Haridas
- A Novel Type of Equipment for Reactive Distillation ModelUploaded byTaylor Penna
- (NATO ASI Series 221) Andrew B. Templeman (Auth.), B. H. v. Topping (Eds.)-Optimization and Artificial Intelligence in Civil and Structural Engineering_ Volume I_ Optimization in Civil and StructuralUploaded byJorge Eduardo Pérez Loaiza
- Piovesan_2011_JBEUploaded byDavide Piovesan
- Optimization_SNOPTUploaded byHamidreza Sarmadi
- DNV Offshore Installation OperationsUploaded byEyoma Etim
- 10.1.1.121Uploaded bygiovannir86
- Cdra Faci Guide (4)_sept 14Uploaded byLinda Himoldang Marcaida
- 2Uploaded byAlejandro Orozco
- Calculation of Sludge Production From Aerobic ASP JanusUploaded byكرم عمرو
- Thesis 1988 OlomolaiyeUploaded byJavier Montenegro
- National Wind Atlas ReportUploaded byFares Haddad
- 00077Uploaded byBehailu Feleke
- HW6_2014Uploaded bySeung Pyo Son
- Effects of Thickening Time on the Application of Cement Slurry for High PressureUploaded byAlexander Decker
- TRUCK EQUIVALENCY FACTORS.pdfUploaded byCathleen Gilmore
- Electrical Arcing Phenomena Based on IEEE 1584 & NFPA 70EUploaded bykspm007
- Risk IdentificationUploaded byNayan Rathod
- Decision Making in MISUploaded byVissen Mooroogen