Pile/Shaft Design Using Artificial Neural Networks (i.e. Genetic Programming)

Pile/Shaft Design Using Artificial Neural
Networks (i.e. Genetic Programming)
with Spatial Variability Considerations
FDOT Contract No.:
Project Manager:
BDK-75-977-68
Rodrigo Herrera, PE
Peter Lai, PE
Principal Investigator:
Michael McVay, PhD
Khiem Tran, PhD
Primary Researcher:
Harald Klammler, PhD
Michael Faraone, EI
Scope
Driven Pile SPT
700
600
Measured [TONS]
• Improvement of
prediction equations (side
and tip resistance) used
by FB-DEEP for both
prestressed concrete piles
and drilled shafts.
• Optimizing prediction
equations by use of a
Genetic Program (GP)
from site insitu and load
test data.
500
400
300
200
100
0
0
100 200 300 400 500 600 700
Predicited [TONS]
2
Research Tasks
1. Data Collection
2. Development of Genetic Code to improve
Pile/Shaft predictions
3. Inclusion of Spatial Variability in assessment
of Pile/Shaft Capacity Equations
4. Development and Evaluation of Pile Capacity
Equations from Insitu Data
5. Final Report and Database Upload
3
Data Collection
• Geotechnical Reports
– SPT Boring Logs
– Lab Test for Rock
•
•
•
•
qu - unconfined compressive strength
qt - split tensile
E - Modulus
REC/RQD – Recovery
• Load Test Reports
–
–
–
–
Drilled Shafts and Driven Piles
Static
Osterberg
Statnamic
4
FDOT Database
• Standardize format
• Organize site data based on hierarchy structure
• Project
• Bridge
• Pier
• Pile/Shaft
• Load Test
• Subsurface
• Borings
• Lab Data
• fdot.ce.ufl.edu
5
Formatting Insitu Data
6
Formatting Load Test Data
7
Driven Pile Datasets
Project Site
Project
Number
Borings (Firm) Load Tests (Firm)
Applied
Foundations
Schmertmann &
Acosta Bridge 72160-3506 Law Engineers
Crapps
5th St. Bascule 412808-1Bridge
52-01
Mactec
Foundation
Pile
Dimension
Length
(Concrete/Ste
(in.)
(ft)
el)
No.
Borings
No. Load
Tests
7
4
18
Concrete
55
53
3
24
Concrete
39-67
Apalachicola
49010-3536
Bay
F.D.O.T.
Schmertmann &
Crapps
28
5
18-24
Concrete
68.2123.7
Apalachicola
49010-3533
River
F.D.O.T.
Schmertmann &
Crapps
33
4
18-30
Concrete
65.293.2
Bayou Chico 48050-3536
F.D.O.T.
Williams Earth
Sciences Inc.
7
3
24
Concrete
46-87
Williams Earth Williams Earth
Blackwater
58002-3449
Sciences Inc.
Sciences Inc.
Bridge (I-10)
4
2
24
Concrete
115
72001-3462
Ardaman &
Associates
Schmertmann &
Crapps
40
8*
30
Concrete
104.5121
Caminida Bay 061-01-0040
Applied
Foundations
LA D.O.T.
4
2
30
Concrete
55-60
35
10
24-30
Concrete
69-125
22
3
24
Concrete
850
Buckman
Bridge
Choctawhatch
60040-3527
ee
230656-1Dixie Highway
52-01
F.D.O.T.
PSI
Schmertmann &
Crapps
Applied
Foundations
Project Site
Project
Number
Foundation
Pile
Dimension
Length
(Concrete/Ste
(in.)
(ft)
el)
30
Concrete 65-110
No.
Borings
No. Load
Tests
6
3*
53
2
24
Concrete
65-92
HDR
5
1
14
Concrete
34.3
HDR
49
4
30
Concrete
52.9101.8
29
3
24
Concrete
80-95
8
4*
24
Concrete
140-150
11
2
18
Concrete
32.834.3
Law Engineers
41
2
30
Concrete
62.5-70
Schmertmann &
Crapps
22
7
20-24
Dames & Moore
19
2
30
Concrete
105-130
Dames & Moore
16
2
24
Concrete
31125.6
Borings (Firm) Load Tests (Firm)
New Driven Pile Datasets Cont.
Dodge Island 87000-3675 Law Engineers
Escambia 48140-3509/
F.D.O.T.
58080-3516
River
Ardaman &
Fort Myers 12001-3513
Associates
Williams and
Howard
15190-3479
Associates
Frankland
Julington
78070-3517
PEC
Creek
Mantanzas
78002-3509
F.D.O.T.
River (SR 312)
Franco/William
Port Orange 79180-3514
s/ Dawson
Roosevelt
89010-3541 Law Engineers
Bridge
Williams and
Sunshine
15170-3421
Associates
Skyway
West Bay
217911-5F.D.O.T.
Bridge
52-01
White City
51020-3514
F.D.O.T.
Bridge
* = includes tension tests
Law Engineers
Schmertmann &
Crapps
Ardaman &
Associates
Williams Earth
Sciences Inc.
Schmertmann &
Crapps
Borings
Totals: 492
Load
Tests
61
Concrete/Stee 38.2l
79.6
9
Drilled Shaft Datasets
Project Site
Project
Number
No.
Borings (Firm) Load Tests (Firm)
Borings
Appalachicola 47010-3519/ Ardaman and
River (S.R.20) 56010-3520 Associates
Law
Fuller Warren 72020-1485
Engineering
Beiswenger,
Gandy Bridge 10130-1544
Hoch & Assoc.
Law
Macarthur
87060-1549
Engineering
Causeway
Williams and
17th Street 86180-1522
Associates
Williams and
Sunshine
15170-3421
Associates
Skyway
Venetian
11120-158Dames &
Causeway
141
Moore
Victory Bridge 53020-3540
Lee Roy
Selmon
10190-1416
F.D.O.T.
PSI
Foundation
No. Load
Dimension Length
Tests
(in.)
(ft)
Schmertmann &
Crapps
148
8
108
80-160
N/A
26
4
36-72
Williams and
Associates
74.5201.5
116
6
48
43.1-83
Law Engineering
44
5
48
30.5150
LOADTEST Inc
95
3
48
40-100
22
10
24-48
38.279.6
17
10
48
50-82
28
6
48
69-100
504
13
48
Borings
Load
Tests
46.779.9
Schmertmann &
Crapps
Florida Testing &
Engineering, Inc
Schmertmann &
Crapps
Applied
Foundation
Totals: 1000
65
10
Data for Genetic Program
60040-3527-Choctawhatchee
2000
60040-3527 Choctawhatchee
20
1500
Load [kip]
0
-20
Pile Name = 38
Diameter = 30 in
Embed Length = 79 ft
1000
500
0
0
-60
10
-100
2
Side = 1050 kips
Davisson Tip = 445 kips
3
-120
-140
-160
0
1.5
1
0.5
Top Displacement [in]
4
-80
Load [kip]
Elevation [ft]
-40
10
Compared to
Telltale Data
2
10
1
20
40
60
80
SPT-N [blows/12 in]
100
10 -2
10
-1
0
10
10
Top Displacement [in]
1
10
11
Genetic Program
• Optimization tool based on
natural selection and
evolutionary algorithms.
• Optimizes a prediction model
based on a set of inputs (insitu
data) and corresponding
outputs (load test).
• Previous work done for driven
pile models using CPT-qt, and
shallow foundation settlement
from SPT-N.
• Begins with generation of
random population of models.
• Model represented by tree
structure
SQ
+
X2
x
7
X1
12
Crossover
Selection of two models from the population
SQ
/
ln
+
x
7
X1
/
SQ
4
X2
ln
X2
4
x
+
-
X2
X1
Resulting new models for next generation
7
X1
X1
X2
13
Mutation
Selection of one model from the population
SQ
SQ
+
/
x
7
Resulting new model for next generation
X2
X1
x
7
X2
X1
14
Reproduction
• One model is selected from the population
and copied over to the next generation.
SQ
SQ
+
+
x
7
x
X2
X1
7
X2
X1
15
Selection Criteria/Fitness
• Need a basis for which models are selected for use in
evolutionary algorithms.
• For symbolic regression can use R2, mean squared error
(MSE), etc. to quantify how well the predict model fits
the measured data.
• Assign a probability of selection for model based on
fitness score relative to entire.
– i.e. Better fitting models have a higher chance of being
selected.
• Repeat this process for multiple generations until
optimal solution is determined.
M – measured value
P – corresponding model prediction
N – number of data points
16
Implementation
and Validation
• Coded in MATLAB using data downloaded
from the FDOT database.
• Validation for solving 1 equation.
5
10
10
x 10
8
15
GP Model
Target Model
6
10
MSE [-]
y [-]
x 10
4
2
5
0
-2
0
50
x [-]
100
0
0
5
10
15
Generation [-]
20
25
17
Validation cont.
• See how well GP predicts
FB-DEEP models.
• Create borings logs of
uniform SPT-N blow
count and soil type.
• Generate multiple profiles
with different SPT-N.
• Use FB-DEEP calculation
to be representative load
test.
SPT-N
Sand
Depth
18
Validation cont.
15
x 10
FB-DEEP SOIL TYPE 1 (Clay) Fitting
a
5
b
3500
18
FB-DEEP MODEL
GP MODEL
x 10
c
10
16
3000
Measured [lbs]
10
5
2500
Minimum MSE [lbs2]
Unit Side Resistance [lbs]
14
2000
1500
12
10
8
6
4
1000
2
0
0
5
10
500
0
15
Predicted [lbs]
x 10
20
40
0
0
60
10
NSPT [lbs]
5
20
30
40
50
40
50
Generation [-]
FB-DEEP SOIL TYPE 3 (Sand) Fitting
x 10
a
b
2500
FB-DEEP MODEL
GP MODEL
9
6
5
4
3
2
1500
1000
3
2.5
2
1.5
1
500
0.5
1
0
0
c
10
3.5
2
7
x 10
4
2000
Unit Side Resistance [lbs]
8
Measured [lbs]
4.5
Minimum MSE [lbs ]
10
5
2
4
6
Predicted [lbs]
8
10
5
0
0
20
40
NSPT [lbs]
60
0
0
10
20
30
Generation [-]
19
Validation cont.
• See how well GP predicts
2 FB-DEEP models at the
same time.
• Create borings logs with 2
uniform layers of SPT-N
blow count and soil type.
• Generate multiple profiles
with different layering
and SPT-N.
• Determine Li, to evaluate
SPT-N’s contribution to
side resistance.
SPT-N
Clay
L1
L2
L3
Sand
L4
L5
Depth
20
GP Capacity Prediction
• GP needs to account for
different layering sizes as
well as sample spacing.
• Li provides means to
determine a weighted
average.
• Driven Pile data set only
provides total resistance,
which side and tip can be
derived from.
• GP then has to evaluate
fitness based on total side
resistance, when optimizing
multiple soil type models.
21
Validation cont.
x 10
a
5
b
10
3500
12
3000
2500
8
2000
x 10
c
10
8
fs [psf]
Measured [lbs]
2
10
FB-DEEP MODEL ST1
FB-DEEP MODEL ST3
GP MODEL ST1
GP MODEL ST3
Minimum MSE [lbs ]
14
6
1500
4
1000
2
500
6
4
2
0
0
5
10
Predicted [lbs]
15
x 10
5
0
0
20
40
NSPT [Blows/12in]
60
0
0
10
20
30
40
50
Generation [-]
GP
Run
1
MSE (lbs2)
2
2.07e+10
3
1.25e+08
1.14e+08
22
GP with Real Data
•
•
•
•
•
9 sites totaling 233 boring 34 load test.
Ultimate side friction from load test data.
Optimizing 2 soil type models
Limited SPT-N to max value of 60
For each site, ultimate side friction is
evaluated for each boring and compared to
the corresponding load test piles.
23
Assignment of Soil Type
• First attempt fitting for
cohesion and
cohesionless models.
• Binning based on either
USCS classification or
boring log soil
description.
• Automatically sorted
using database fields.
• Advantages:
– less computation time
– Learn how GP works with
real data
Soil Type 1
CL
ML
CL-ML
CH
MH
OL
OH
SC
SW-SC
SP-SC
SM-SC
SW-SM
SP-SM
Pt
Soil Type 3
GW
GP
GM
GC
GC-GM
GW-GM
GW-GC
GP-GM
GP-GC
SW
SP
SM
24
Results
550
3500
GP ST1
GP ST3
FB-DEEP ST1
FB-DEEP ST3
500
3000
Unit Side Friction [psf]
450
Measured [tons]
400
350
300
250
200
2500
2000
1500
1000
150
500
100
50
0
200
800
600
400
GP Prediction [tons]
1000
1200
0
10
40
30
20
SPT-N [blows/ 12in]
50
60
Large scatter can partially be to GP’s model error and spatial
variation across the site.
25
Next Phase of GP Work
• Entering Additional FDOT sites into Online
database
• Investigation of Different Soil/Rock descriptors
– Refining of USCS binning.
• Task 3, inclusion of spatial variability in error
assessment.
• Investigate the separation of site into multiple
zones due to stratigraphy or anisotropy
26
Next Phase of GP Work -cont
• Evaluate Tip resistance based on soil
&rock type.
• Develops Models based on Davisson and
Ultimate
• Evaluate Skin and Tip Resistance for
Drilled shafts using GP
27
Questions?
28