Soft Computing Approaches to Constitutive Modeling in Geotechnical Engineering

Transcription

1 Soft Computing Approaches to Constitutive Modeling in Geotechnical Engineering Jamshid Ghaboussi and Youssef M. Hashash Department of Civil and Environmental Engineering University of Illinois at Urbana-Champaign

2 Contents Introductory Remarks on Soft Computing SelfSim: Constitutive Modeling from System Response Several Applications of SelfSim Field Calibration of constitutive Models in Deep Excavations

3 Soft Computing Soft computing methods are inspired by the computing and problem solving strategies in nature. They Are fundamentally different than the mathematically based problem solving methods. Soft computing tools are neural networks, genetic algorithm and fuzzy logic.

4 Problem Solving in Nature Nature solves very difficult inverse problems. Nature does not use mathematics. Two main components of nature s problem solving strategies are: Learning Reduction in disorder Nature utilizes random search and gradual improvement.

5 Forward Problems

6 Inverse Problems Type 1. Input is unknown Type 2. System is unknown

7 Inverse Problem of Constitutive Modeling Material Test

8 Inverse Problem of Constitutive Modeling from Field Measurements In many geotechnical problems it is too difficult to sample or perform realistic material tests. Often field measurements are available.

9 Autoprogressive Algorithm

10 Autoprogressive Algorithm Structure is modeled by FE and constitutive model is represented by a NN. Pre-train NN material model to learn linearly elastic constitutive behavior Apply boundary forces: apply P n, compute U n P = P + P n n 1 n KδU = P I j j 1 t n n n U = U + δu j j 1 j n n n σn = σnnn( εn,... :...) Enforce boundary displacements Train Neural Networks δun = Un + Un ( σ + δσ ) = ΝΝ(( ε + δε ),... :...) n n n n σ = σ NN(( ε + δε ),... :...) n n n n

11 Learning Material Behavior from Response of a Simple Structure FE Analysis A, Force BC NN Training FE Analysis B, Displacement BC σ x σ y τ xy ε ε x y γ xy Simulated Experiment, Von-Mises Plasticity with hardening Train a NN to learn the material behavior from the response of this structure

12 Evolution of Load-Displacement pre-training pass load (MN) edge-center point deflection (mm)

13 NN Learning of Material Behavior s11-e11 (pre-train) (pass 4) ) 8) 1) 11) 12) 13) s22-e22 (pre-train) (pass 4) ) 8) 1) 11) 12) 13) 3.E+2 2.E+2 2.E+2 1.E+2 1.E+2.E+1.E+ -6.E-3 -.E-3-4.E-3-3.E-3-2.E-3-1.E-3.E+ 9.E+2 8.E+2 7.E+2 6.E+2.E+2 4.E+2 3.E+2 2.E+2 1.E+2.E+.E+ 1.E-2 2.E-2 3.E-2 4.E-2.E-2 6.E-2 7.E-2

14 NN Learning of Material Behavior s22-e22 (pre-train) (pass 4) ) 8) 1) 11) 12) 13) s12-e12 (pre-train) (pass 4) ) 8) 1) 11) 12) 13) 9.E+2 8.E+2 7.E+2 6.E+2.E+2 4.E+2 3.E+2 2.E+2 1.E+2.E+.E+.E-3 1.E-2 1.E-2 2.E-2 2.E-2 3.E-2 7.E+1 6.E+1.E+1 4.E+1 3.E+1 2.E+1 1.E+1.E+ -1.E-3.E+ 1.E-3 2.E-3 3.E-3 4.E-3.E-3 6.E-3-1.E+1-2.E+1

15 Time dependent behavior of reinforced concrete

16 Creep of Concrete Beam 1 mm C L Experiment by Bakoss et al. (1982) mm 2 mm 2 mm deflection (mm) mm pass pass pass1 pass time (days) top surface instant 12 days 6 days 13 days 13 days bottom pass pass1 pass2 pass strain surface..1.2

17 Field Calibration of Time- Dependent Behavior in Segmental Bridges Using Self-Learning Simulation

18 Case study: Pipiral Bridge (Colombia) each span: 12m

19 Jan 1, 22: 73 days (12 segments) Feb 14, 22: 18 days (all 18 segments) Jul 11, 22: 2 days (all 18 segments) design camber at the completion The Challenge: Deflection Control camber (mm) distance from the pier P4 (m)

20 Field Calibration of the Pipiral Bridge displacement (mm) step -2 pre-train pass 12-2 measurement distance from the pier (m) displacement (mm) step distance from the pier (m)

21 Improvement of Camber by Using SelfSim adjustment to casting curves (mm) camber (mm) Jan 1, 22: 73 days (12 segments) Feb 14, 22: 18 days (all 18 segments) Jul 11, 22: 2 days (all 18 segments) design camber at the completion distance from the pier (m) distance from the pier P4 (m)

22 Learning the hysteretic behavior of soil from the seismic site response

23 Application of SelfSim to synthetically generated downhole array data- case Hyperbolic model + Masing rule(target) Initialized NN (Linear) Hyperbolic model + Masing rule(target) Initialized NN (Linear) SelfSim (1st pass) Displacement(ft) Displacement(ft) Layer Sinusoidal Sinusoidal motion motion single single layer layer Time (sec) Hyperbolic model + Masing rule(target) Initialized NN (Linear) SelfSim (3rd pass) Time (sec) Hyperbolic model + Masing rule(target) Initialized NN (Linear) SelfSim (7th pass) Displacement(ft) Displacement(ft) Time (sec) Time (sec)

24 Application of SelfSim to synthetically generated downhole array data- case Hyperbolic model + Masing rule (Target) Initialized NN (Linear) Hyperbolic model + Masing rule (Target) Initialized NN (Linear) SelfSim (1st pass).2.2 Stress. -.2 Stress Layer1 cyclic cyclic motion motion single single layer layer Strain Hyperbolic model + Masing rule (Target) Initialized NN (Linear) SelfSim (3rd pass) Strain Hyperbolic model + Masing rule (Target) Initialized NN (Linear) SelfSim (7th pass) Stress. -.2 Stress Strain Strain

25 Application of SelfSim to synthetically generated downhole array data- case2 1 Pass1 1 Pass DEESPSOIL (NN) DEEPSOIL (hyperbolic) DEEPSOIL+LinearNN DEESPSOIL (NN) DEEPSOIL (hyperbolic) DEEPSOIL+LinearNN 6 6 Layer1 Sa (g) 4 Sa (g) 4 Layer Layer3 seismic seismic motion motion multiple multiple layers layers same same soil soil properties properties Layer Period (sec) Pass Period (sec) 1 Pass DEESPSOIL (NN) DEEPSOIL (hyperbolic) DEEPSOIL+LinearNN DEESPSOIL (NN) DEEPSOIL (hyperbolic) DEEPSOIL+LinearNN 6 6 Sa (g) 4 Sa (g) Period (sec) Period (sec)

26 Application of SelfSim to synthetically generated downhole array data- case2 4.E-2 3.E-2 Layer1 Layer2 Displacement(ft) 2.E-2 1.E-2.E+ -1.E-2 Layer3 seismic seismic motion motion multiple multiple layers layers same same soil soil properties properties Layer4 Stress -2.E-2 SelfSim -3.E-2 DEEPSOIL+hyperbolic(Target) DEEPSOIL+linearNN -4.E Target Layer4 Layer3 Layer2 3.E+3 2.E+3 1.E+3 Tim e (sec) Layer1.E+ -3.E-1-2.E-1-1.E-1.E+ 1.E-1 2.E-1 3.E-1-1.E+3-2.E+3-3.E+3-4.E+3 -.E+3 Strain(%)

27 SelfSim in Bio-Engineering Applications in Ophthalmology J. Ghaboussi, D. A. Pecknold, Y. M. Hashash Characterizing the material properties of human cornea for simulating Individualized laser surgery Measurement of intra-ocular pressure (IOP) Developing a new instrument Applications in bio-medical imaging Characterizing the material properties of soft tissue Imaging of the material properties of soft tissue for diagnostics and disease progression

28 Simulation of Goldman Applanation Tonometer.2mm Applanation mmhg Before loading steps First step Second step

29 Real-time soil model for toolmedium interaction in virtual reality environment

30 DEM Machine-Medium Interaction

31 DEM Machine-Medium Interaction

32 DEM Machine-Medium Interaction

33 Integration of numerical modeling and field observations for deep excavations

34 1. Field Measurements Initializing or Pre-training stress strain data from: 1. Linear elastic 2. Laboratory tests 3. Case histories 4. Approximate constitutive models σ,ε 2. SelfSim Training FEM, iterated σ Stress-Strain Pairs Training of NANN ε a) Simulate Construction Sequence => extract stresses Neural Network Constitutive Soil Model b) Apply Measurements => extract strains 3.Forward FEM analysis with trained NN material model and/or Next excavation stage Similar excavations

35 Synthetic Field Measurements B/2=2m H=m Support Spacing =2.m Diaphragm Wall Thickness =.9m E=2.3*1 4 MPa υ=.2 Depth (m) Lateral Wall Deflection (cm) MIT-E3 "Field" Measurements Surface Settlement (cm) Distance behind the wall (m) MIT-E3 "Field" Measurements -4 -

36 No SelfSim training Depth (m) Lateral Wall Deflection (cm) Elastic Pre-trained Surface Settlement (cm) Distance behind the wall (m) Elastic Pre-trained -4 -

37 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #1 using Exc. depth =2.m data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #1 using all Exc. depth =2.m data

38 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #1 using Exc. depth=m data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #1 using Exc. depth=m data

39 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #1 using Exc. depth=7.m data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #1 using Exc. depth=7.m data

40 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #1 using Exc. depth=1m data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #1 using Exc. depth=1m data

41 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #1 using Exc.depth=12.m data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #1 using Exc.depth=12.m data

42 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #1 using all Exc. data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #1 using all Exc. data

43 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #2 using all Exc. data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #2 using all Exc. data

44 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #4 using all Exc. data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #4 using all Exc. data

45 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #6 using all Exc. data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #6 using all Exc. data

46 SelfSim training Depth (m) Lateral Wall Deflection (cm) Training: Pass #8 using all Exc. data Surface Settlement (cm) Distance behind the wall (m) Training: Pass #8 using all Exc. data

47 Evaluate a Trained NN model H CL Support Spacing h=2.m Distance from Diaphragm Wall, x (m) σ L SelfSim Analysis Diaphragm Wall Length, L=4m Thickness, t =.9m E = 2.3x1 4 MPa ν =.2 ε D th ( ) Stress and Strain σ j ε j ε j 1 σ j 1 Simulate Laboratory Tests H Surface Displacement (mm) Depth (m) CL Support Spacing h=2.m Distance from Diaphragm Wall, x (m) L 22. Trained NN model Diaphragm Wall Length, L=4m Thickness, t =.9m E = 2.3x1 4 MPa ν = D th ( ) Distance behind wall (m) NN. MIT E mm

48 Plane Strain Tests (PST) (σ' v -σ' h )/2*σ' v PST Extension PST compression ε (%) axial (σ' v -σ' h )/2*σ' v ε (%) axial (σ' v -σ' h )/2*σ' v ε (%) axial SelfSim MITE3 OCR1. SelfSim Elastic Pre-train 1 SelfSim MITE3 OCR1. SelfSim Elastic Pre-train SelfSim MITE3 OCR1. SelfSim Elastic Pre-train Secant Modulus, (δσ -δσ )/2γσ' MITE3 SelfSim Maximum Shear Strain, γ (%)

49 Learning from Precedence h e= 2.m h s=.m B/2=4m Case3 h e= 2.m h s= 2.m B/2=2m 4 m Flexible Wall m 3 m 4 m 4 m New Excavation Site 4 m Stiff Wall 3 m 2 m 6 m Crust Clay Layer MIT-E3 OCR=4 Soft Clay Layer MIT-E3 OCR=1.3 Stiff Layer MIT-E3 OCR= Rock Linear Elastic h e= 2.m h s= 2.&. m 3 m Stiff Wall Case2 m 3 m h e= 2.m h s= 2.m B/2=2m 4 m Flexible Wall Case1 m 4 m 3 m 4 m B/2=3m 4 m 4 m

50 Depth (m) Case1 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Flexible Wall h s =2.m B/2=2m 12. Case 1 after training with all three cases Soil behavior to be learned SelfSim Training Surface Settlement (cm) Depth (m) Case2 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Stiff Wall h s =. & 2.m B/2=3m Case2 after training with all three cases Soil behavior to be learned Surface Settlement (cm) Case3 Distance behind the wall (m) 1 2 Depth (m) Lateral Wall Deflection (cm) h =2. &.m t B/2=4m m Flexible Wall 12. Exc. Depth (m) Case 3 after training with all three cases Soil behavior to be learned Surface Settlement (cm)

51 Depth (m) Case1 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Flexible Wall h s =2.m B/2=2m 12. Case 1 after training with all three cases Soil behavior to be learned SelfSim Training Surface Settlement (cm) Depth (m) Case2 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Stiff Wall h s =. & 2.m B/2=3m Case2 after training with all three cases Soil behavior to be learned Surface Settlement (cm) Case3 Distance behind the wall (m) 1 2 Depth (m) Lateral Wall Deflection (cm) h =2. &.m t B/2=4m m Flexible Wall 12. Exc. Depth (m) Case 3 after training with all three cases Soil behavior to be learned Surface Settlement (cm)

52 Depth (m) Case1 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Flexible Wall h s =2.m B/2=2m 12. Case 1 after training with all three cases Soil behavior to be learned SelfSim Training Surface Settlement (cm) Depth (m) Case2 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Stiff Wall h s =. & 2.m B/2=3m Case2 after training with all three cases Soil behavior to be learned Surface Settlement (cm) Case3 Distance behind the wall (m) 1 2 Depth (m) Lateral Wall Deflection (cm) h =2. &.m t B/2=4m m Flexible Wall 12. Exc. Depth (m) Case 3 after training with all three cases Soil behavior to be learned Surface Settlement (cm)

53 Depth (m) Case1 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Flexible Wall h s =2.m B/2=2m 12. Case 1 after training with all three cases Soil behavior to be learned SelfSim Training Surface Settlement (cm) Depth (m) Case2 Distance behind the wall (m) 1 2 Lateral Wall Deflection (cm) m Stiff Wall h s =. & 2.m B/2=3m Case2 after training with all three cases Soil behavior to be learned Surface Settlement (cm) Case3 Distance behind the wall (m) 1 2 Depth (m) Lateral Wall Deflection (cm) h =2. &.m t B/2=4m m Flexible Wall 12. Exc. Depth (m) Case 3 after training with all three cases Soil behavior to be learned Surface Settlement (cm)

54 Prediction of New Case Depth (m) Lateral Wall Deflection (cm) h s =2.m B/2=2m Soil behavior to be learned Distance behind the wall (m) Surface Settlement (cm) 4

55 Prediction of New Case after Case1 Depth (m) Lateral Wall Deflection (cm) Prediction of "New" Case h s =2.m B/2=2m Soil behavior to be learned Distance behind the wall (m) 1 2 Prediction of "New" Case after training with Case Surface Settlement (cm)

56 Prediction of New Case after Case1&2 Depth (m) Lateral Wall Deflection (cm) Prediction of "New" Case h s =2.m B/2=2m Soil behavior to be learned Distance behind the wall (m) 1 2 Prediction of "New" Case after training with Case 1 and Surface Settlement (cm)

57 Prediction of New Case after Depth (m) Lateral Wall Deflection (cm) Prediction of "New" Case h s =2.m B/2=2m Case1,2&3. 2. Soil behavior to be learned Distance behind the wall (m) Prediction of "New" Case after training with all three cases Surface Settlement (cm)

58 LURIE RESEARCH CENTER

59 Lurie Research Center Plan and Cross Section View N. Fairbanks Court LR4 Pedestrian tunnel LR LR6 LR7 LR3 E. Superior Street Lurie Research Center LR8 LR2 E. Huron Street LR1 Prentice Pavillion (1level basement) N PK Nail Embedded Soil Settlement Point Utility Monitoring Point Inclinometer Soil Boring El m El m El m El. -.2 m El m PZ27 Sheeting El m Lake Sand Fill Soft/Med Silty Clay Stiff/V. Stiff Silty Clay Hard Silty Clay El. +.6 m El m El m El m feet 6 meters

60 Lat. Wall Deflection (mm) Elevation (m) (a) Stage 1: Exc. +1. m (g) Stage 7: Exc m Elevation (m)

61 Simplified Construction Sequence C L +4.2 C L Reference Configuration Stage Undrained Excavation 16 days dewatering Stage 1 C L C L Installation Tieback 1 Undrained Excavation Stage 2 86 days dewatering Stage 3 C L C L Installation Undrained Excavation Tieback 2 38 days dewatering Stage 4 Stage C L +4.2 C L Installation Undrained Excavation Tieback 3 Stage 6 2 days dewatering Stage 7

62 No Training Elevation (m) Lateral Wall Deflection (mm) (a) No Training Distance behind the wall (m) Exc Stage Exc. 1. m Tieback 1 Exc m Tieback 2 Exc. -.8 m Tieback 3 Exc m Measured (b) No Training Pre-trained Surface Settlement (mm)

63 SelfSim Training Elevation (m) Lateral Wall Deflection (mm) (a) 1 Passes of Training Distance behind the wall (m) Exc Stage Exc. 1. m Tieback 1 Exc m Tieback 2 Exc. -.8 m Tieback 3 Exc m (b) 1 Passes of Training Measured SelfSim Surface Settlement (mm)

64 Lat. Wall Deflection (mm) SelfSim Training Lat. Wall Deflection (mm) Lat. Wall Deflection (mm) 1 - Elevation (m) (a) Stage 1: Exc. +1. m (b) Stage 2: Tieback 1 (c) Stage 3: Exc m Elevation (m) Elevation (m) Elevation (m) (d) Stage 4: Tieback (f) Stage 6: Tieback 3 2 Lurie Center LR3 LR LR6 LR7 LR8 Representative SelfSim (e) Stage : Exc. -.8 m (g) Stage 7: Exc m Elevation (m) Elevation (m)

65 Surface Settl. (mm) Surface Settl. (mm) Surface Settl. (mm) Surface Settl. (mm) Distance behind the wall (m) SelfSim Training (a) Stage 1: Exc. +1. m (c) Stage 3: Exc m (e) Stage : Exc. -.8 m (g) Stage 7: Exc m Distance behind the wall (m) (b) Stage 2: Tieback (d) Stage 4: Tieback (f) Stage 6: Tieback Surface Settl. (mm) Surface Settl. (mm) Surface Settl. (mm)

66 Assessment of extracted soil model.6.4 (σ' v -σ' h )/2σ' v ε (%) axial Extracted Clay Model (SelfSim) Triaxial tests (Holman, 2) Elastic Pre-training

67 Assessment of extracted soil model.6.4 (σ' v -σ' h )/2σ' v ε (%) axial Extracted Clay Model (SelfSim) Triaxial tests (Holman, 2) Elastic Pre-training

68 Concluding Remarks Most of the difficult engineering problems are inverse problems. Very effective methods can be developed for solving these inverse problems by learning from nature s problem solving strategies. Neural networks, genetic algorithm and fuzzy logic enable development of these methods. Soft computing methods use nature s strategies of reduction in disorder and learning.

69 Concluding Remarks Soft computing methods are fundamentally different from the conventional mathematically based methods. Soft computing methods operate with Imprecision tolerance Non-universality Functional non-uniqueness.