Evolution of Geotechnical Reliability

John T. Christian
Consulting Engineer Prof. Civil & Env. Engg.
Burlington, Massachusetts Univ. Mass. – Lowell
[email protected] [email protected]
781-272-3196 978-934-6763

2019 CGS – SOS Workshop

Fundamentals of Risk in Geotechnical Engineering
Evolution of Geotechnical Reliability
• Background and early developments
• Basic principles of reliability and risk analysis
• Difficulties and misconceptions
• Bayesian updating
• Communicating risk and relation to the observational method
• TONEN project example
• Conclusions and expectations
Evolution of Geotechnical Reliability
Background and early developments
Uncertainty and risk (or lack of
reliability) have been part of the
human condition for a long time.
The geotechnical engineering
community has been relatively late
in addressing them rationally, but
much has happened in recent
years.

22 October, 1895,
Gare Montparnasse,
Paris
Earlier Contributors
• Gregory B. Baecher
• C. Allin Cornell
• Herbert Einstein
• Milton E. Harr
• Fred D. Kulhawy
• Peter Lumb
• Wilson H. Tang
• Erik Vanmarcke
• Daniele Veneziano
• Steven G. Vick
• T. H. Wu
Some References
• A. Casagrande (1964) “Role of the ‘Calculated Risk’ in
Earthwork and Foundation Engineering”
• R. V. Whitman (1981) “Evaluating Calculated Risk in
Geotechnical Engineering”
• S. G. Vick (2002) Degrees of Belief
• G. B. Baecher & J. T. Christian (2003) Reliability and
Statistics in Geotechnical Engineering
• G. A. Fenton & D. V. Griffiths (2008) Risk Analysis in
Geotechnical Engineering
Evolution of Geotechnical Reliability

Basic principles of
reliability and risk
analysis
“We ought to fear or hope for an event
not only in proportion to the advantage
or disadvantage but also with some
consideration of the likelihood of the
occurrence.”
Blaise Pascal, quoted by Ian Hacking, The Emergence
of Probability, Cambridge University Press, 1975.
How to deal with uncertainty in geotechnical
engineering?
How to deal with uncertainty in geotechnical
engineering?
• Ignore it
How to deal with uncertainty in geotechnical
engineering?
• Ignore it

• Be Conservative
How to deal with uncertainty in geotechnical
engineering?
• Ignore it

• Be Conservative

• Use the observational method (“Adaptive Management”)

How to deal with uncertainty in geotechnical
engineering?
• Ignore it

• Be Conservative

• Use the observational method

• Quantify it
 Reliability

M=R-Q

Define β = µM / σM
Reliability

μM = μR -μQ
σM2 = σR 2+ σQ2
β = µ M / σM
pf= Φ[-β]
 Reliability
We can also work with the Factor of Safety:
F=R/Q

But then it is easier to use the logarithms:

ln F = lnR – ln Q

The rest of the mathematics is the same as for the

Margin of Safety except that the expression for σM
becomes messier.
 Requirements
• Input

• Methodology for reliability analysis

• Geotechnical analytical model

• Output and interpretation

Evolution of Geotechnical Reliability

Difficulties and
misconceptions
 Nature of Uncertainty
“. . . the laws of probability, so true in general,
so fallacious in particular, . . .”

Edward Gibbon
 Nature of Uncertainty
Is the state of nature really random?
i. e. aleatory

Or is it fixed, but we simply do not know what it

is?
i. e. epistemic
 The Meaning of Probability
Is probability a measure of the frequency with
which things happen?

Or is it a measure of our degree of belief in a

particular state of affairs?

Or is it some combination of these?

 The Meaning of Probability

The insurance company is a frequentist when it

sells you insurance.

But . . .
 The Meaning of Probability

The insurance company is a frequentist when it

sells you insurance.

But you buy insurance based on your degree of

belief.
 The Meaning of Probability

When processing laboratory data from many

tests, we are acting like frequentists.

When carrying out an exploration program, we

are trying to sharpen our degree of belief.
 Measurements of Vs
Six values measured:
229, 224, 229, 217, 200, 241 m/s

These give µ ≈ 223 m/s

σ ≈ 13.9 m/s
 Measurements of Vs
Six values measured:
229, 224, 229, 217, 200, 241 m/s

These give µ ≈ 223 m/s

σ ≈ 13.9 m/s

Actual values: µ = 240 m/s

σ = 24 m/s
 Measurements of Vs

Actual
Estimated

150 200 250 300 350

Vs (m/s)
 The 6-σ method

Since 99.7% of all the values of a Normally distributed variable

fall between μ+3σ and μ -3σ, why not simply take the
difference between the largest and smallest values we can
have ever seen and divide by 6 to get an estimate of σ?
5

0
0 10 20 30 40 50
Number of Points in Sample

Conclusion: Do not use 6σ.

Evolution of Geotechnical Reliability

Bayesian updating
Frequentist versus Bayesian Statistics
Frequentist statistics essentially give p[data|state of nature]

Bayesian statistics essentially give p[state of nature|data]

 An exploration program in a sandy deposit.
• What is the probability that a liquefiable zone exists?

• How is this probability affected by the results of successive

borings?

• Are more borings justified?

 Conditional Probabilities
The probability of finding a liquefiable zone in any one
boring is about 0.3. Also, it is possible to get a false
positive when no liquefiable material exists. Say that this
has about 0.05 probability. So:

=P F E  0.3
= P F E  0.7
=P F E  0.05
= P F E  0.95
   
 Results for P0 = 0.5
If one boring gives a positive result,
Suppose that existence and non-existence of
liquefiable material are equally likely, so P0[E] = 0.5.

P F E  P0 E 
P1 E F  =
P F E  P0 E  + P F E  P0 E 
 
0.3 0.5 
= 0.86
0.3 0.5  + 0.05  0.5 
 Results for P0 = 0.5
Under the same conditions, if one boring gives a
negative result,

P F E  P0 E 
P1 E F  =
  P F E  P E  + P F E  P E 
  0    0 
0.7  0.5 
= 0.42
0.95  0.5  + 0.7  0.5 
 For three trials with P0=0.5
Bayesian Updating of Results
F
1
F not F
F
0.8 not F
not F
F
F
0.6
not F F
not F
0.4
not F

0.2

0
0 1 2 3
Trial Number
 Effect of Initial Estimate of P[E]

0.8
P[E] after 3 Trials

0.6 3 hits
2 hits
1 hit
0 hits
0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Initial Estimate of P[E]
Evolution of Geotechnical Reliability

Communicating
risk and relation
to the
observational
method
 Recognizing Risk
People simply do not realize how often failures happen.
 Recognizing Risk
People simply do not realize how often failures happen.
For example, Baecher, Paté-Cornell, and de Neufville
studied records of dam failures and concluded that the
annual probability of failure of an engineered dam is 10-4
or one failure per 10,000 dam-years.
 Recognizing Risk
People simply do not realize how often failures happen.
For example, Baecher, Paté-Cornell, and de Neufville
studied records of dam failures and concluded that the
annual probability of failure of an engineered dam is 10-4
or one failure per 10,000 dam-years.
FEMA indicates that there are > 83,000 “large” dams in
the US. “Large” is defined as > 25 ft. high and > 50 acre-
feet of reservoir or otherwise critical.
 Recognizing Risk

National Performance of Dams Project, Stanford University

Estimating
Overconfidence Bias

Probability
1

after Vick (2002) 0.1

Actual Error Frequency

0.01

0.001

0.0001

0.00001

0.000001
0.000001 0.00001 0.0001 0.001 0.01 0.1 1
Subjective Probability of Error
 Expert Elicitation
A 40 ft. high (approx.) embankment had been built as part of a planned
highway I-95. When the highway project was cancelled, the
embankment was left in place.
A panel of internationally recognized geotechnical experts was asked to
predict how much additional fill had to be placed on the embankment
to bring it to failure. They were provided with extensive field and
laboratory data.
The embankment was raised to failure. The predictions and
observations are compared on the next slide.
 Expert Elicitation
(after Hynes & Vanmarcke 1975)
I-95 Expert Estimations

35

30
predicted additional fill (ft)

25
actual fill added at
failure = 18.7 ft.
20

15
average best
estimate = 15.6 ft.
10

1 2 3 4 5 6 7
0
expert
 95% Confidence Bounds for Compression Ratio
Another San Francisco Bay Mud
0.45
example:

Experts were 0.4

asked to
estimate the 0.35
confidence
bounds for 0.3
the San
Francisco Bay
0.25
Mud CR.
0.2
 Communicating Risk
Calculated probability is at best a lower bound.

People have a lot of trouble with small probabilities.

What is the allowable probability of failure?

 Communicating Risk

The F-N diagram –

Baecher’s original
version
 Communicating Risk
A straight line on an F-N plot means that the product of the
probability of failure and its consequences is constant.

This implies that an event with low consequences and high

probability is equal to another event with high consequences
and low probability.

This may not be what we want.

 F/N Plot
Several versions:
10-2 10-2

Unacceptable

Frequency of accidents with N or more fatalities

10-3
Frequency of accidents with N or more fatalities

10-3
Unacceptable 10-3

Probability of at least N fatalities

Limit
10-4
10-4
10-4

10-5
10-5 unacceptable
10-5 ALARP
ALARP
10-6
10-6 Very high
consequence dam
reduction De minimis region
10-6
10-7
desired Objectiv e
10-7
Acceptable Acceptable
10-8 10-7
10-8 acceptable

10-9
10-8
10-9 0
1 10 100 1000 10,000
10 101 102 103 0 1 10 100 1000
Number of Fatalities, N
Number of Fatalities, N
Number of prompt fatalities, N
Hong Kong Planning Netherlands Risk Criterion ANCOLD
Department
 F-N Estimates for California and NOLA Levee Systems
1.E+00

ANCOLD Tolerable Risk for

Existing

1.E-01 California Data

NOLA Results
Annual Exceedance Probability

1.E-02

1.E-03

1.E-04

1.E-05

1.E-06
1 10 100 1,000 10,000 100,000
Fatalities
 Communicating Risk
There other useful results we can get from reliability analyses:

Comparative probabilities of “failure”

Contributions of different components

 Normally Distributed FS
1

We can reduce 0.1

the probability of
failure by
SD = 0.1

Probability of Failure
SD = 0.2

(a) increasing the SD = 0.3

SD = 0.4

mean FS or
0.01
SD = 0.5
SD = 0.6

(b) reducing its SD = 0.7

SD = 0.8

standard SD = 0.9
SD = 1.0

distribution
0.001

(uncertainty).
0.0001
1 1.2 1.4 1.6 1.8 2
Factor of Safety
 LogNormally Distributed FS
1

We can reduce the

probability of 0.1

failure by SD = 0.1
SD = 0.2

(a) increasing the

Probability of Failure
SD = 0.3
SD = 0.4

mean FS or 0.01
SD = 0.5
SD = 0.6

(b) reducing its SD = 0.7

SD = 0.8

standard SD = 0.9
SD = 1.0

distribution 0.001

(uncertainty).
0.0001
1 1.2 1.4 1.6 1.8 2
Factor of Safety
 Example – Prescribed Criteria
Seismic Min. μFS Max. pf
Case
Factor
0.00 1.2 20%
Inter-ramp 0.07 1.1 40%
0.16 1.0 50%
0.00 1.3 5%
Global 0.07 1.2 20%
0.16 1.0 40%
 Example – Prescribed Criteria
FScrit
Mean Factor of Safety, FS

pf < pfcrit pf < pfcrit

FS < FScrit FS > FScrit

Redesign to increase Design is

mean FS satisfactory

pfcrit
pf > pfcrit pf > pfcrit
Failure Probability, pf

FS < FScrit FS > FScrit

Need both better Need better

engineering and engineering to
redesign reduce uncertainty
 Comparative Probabilities of Failure

Present Design Target

Case
E[FS] pf pf FS
H=6m 1.58 2.5 x 10-2 0.01 1.63
H = 12 m 1.53 4.7 x 10-3 0.001 1.50
H = 23 m 1.50 7.1 x 10-4 0.0001 1.43
Contributions of Different Components
Evolution of Geotechnical Reliability

An example
 Decrease in Risk by Dewatering

Annual Proabability of Off-Site Spill, %

KSS Site
Site 200
Site 400

0
0 1 2 3 4 5 6 7
Amount of Dewatering, m
 Decrease in Risk by Uniform Dewatering

1.E+01

Annual Probability of Off-Site Spill, %

1.E+00

KSS Site
Site 200
Site 400

1.E-01

1.E-02
0 1 2 3 4 5 6 7
Amount of Dewatering, m
 Results
Management decided it was cost-effective to build a slurry
wall around each tank farm and to install a permanent
dewatering system.
 Results
Management decided it was cost-effective to build a slurry
wall around each tank farm and to install a permanent
dewatering system.
This became a standard for other tank farms in the area.
Although other facilities in the area suffered considerably
during the 2011 Tohoku earthquake (Mw = 9.0), this facility
rode through the event with no damage.
 Results
Management decided it was cost-effective to build a slurry
wall around each tank farm and to install a permanent
dewatering system.
This became a standard for other tank farms in the area.
Evolution of Geotechnical Reliability

Conclusions and
expectations
 Conclusions and Expectations

Failures occur for many reasons:

• Pushing the technology too far
• Ignoring prior lessons
• Administrative confusion
• Decisions under pressure
 Conclusions and Expectations

Reliability and probabilistic methods

• Force us to recognize uncertainty
• Quantify the effects of uncertainty
• Place risk in context
• Identify the most effective areas for
improvement
 Conclusions and Expectations

Uncertainty is a central part of geotechnical engineering.

Reliability and risk analysis provide one way to address it

rationally.

In recent decades the profession has made great

progress in understanding risk and reliability.
 Conclusions and Expectations

As always, there are subtleties and unresolved issues.

Successful application of reliability and risk based

methods requires a decision maker – as does the
observational method.

Who owns the uncertainty?

Thank you.