Reliability & Sensitivity Analysis

using quoFEM
Sang-ri Yi
University of California, Berkeley

April 21 2022
NSF award: CMMI 1612843
 NHERI
Virtual
experimental
facility

Sang-ri Yi yisangri@berkeley.edu 2
 SimCenter
Develop Opensource software tools for researchers in Natural hazard engineering

Run simulation

Researchers
Industry
Government Agencies

Sang-ri Yi yisangri@berkeley.edu 3
 SimCenter Tools
UQ-enabling Tool

Deterministic model

UQ/Optimization
Analysis
Random Variables
E A F

Sang-ri Yi yisangri@berkeley.edu 4
 SimCenter Tools
Hazard-specific Modeling and Analysis Tools

Hazard models Structure models UQ analysis

Your own model Your own model

+ Damage and
or hazard scenario generator or building model generator
Loss Analysis

Sang-ri Yi yisangri@berkeley.edu 5
 SimCenter Tools
Regional-Scale Risk Management Tool

Multi-hazard Multi-asset
Simulation Analysis
Earthquakes Buildings

e.g. “The buildings are mostly likely in moderate damage

states. The non-structural damage would dominate the
economic losses. The repair costs range from 1% to 7%
of the total replacement costs, and the repair time range
from 1 to 20 days.”

Hurricanes Lifeline Networks

Sang-ri Yi yisangri@berkeley.edu 6
 SimCenter Tools
ASSET
CHARACTERIZATION We are running UQ workflow
UNCERTAINTY QUANTIFICATION

HAZARD RV1
DESCRIPTION

SUPPORTING DATABASES
QoI1 Mean
Std.Deviation
ASSET & EVENT RV2 Kurtosis
REPRESENTATION Computer
Skewness
Simulation Probability of failure
RESPONSE QoIM
ESTIMATION Sensitivity
RVN QoI = f (RV1, RV2, ..RVN) PDF, CDF
PERFORMANCE …
ASSESSMENT Efficient UQ Algorithms + HPC

RECOVERY
SIMULATION

SimCenter’s scientific workflow

Sang-ri Yi yisangri@berkeley.edu 7
 Introduction to quoFEM
quoFEM:
Quantified Uncertainty with Optimization for the Finite Element Method
ASSET
CHARACTERIZATION EE-UQ, WE-UQ, Hydro-UQ, PBE, R2D
- Helps generate event scenarios for specified hazards
UNCERTAINTY QUANTIFICATION

HAZARD SUPPORTING DATABASES

DESCRIPTION
User- and provides seamless workflow
provided
ASSET & EVENT
REPRESENTATION
simulation quoFEM
RESPONSE
model
ESTIMATION - Flexible to problem/model types
PERFORMANCE - Strong UQ capacity
ASSESSMENT

RECOVERY
SIMULATION

Sang-ri Yi yisangri@berkeley.edu 8
 quoFEM (v.3.0)
Algorithms
FEM Engines Methods • Latin Hypercube Sampling
• Monte Carlo
Sampling • Gaussian Process Regression
OpenSees • Polynomial Chaos Expansion
Orange: Dakota engine
Global • Probability model-based approximation
(Sandia National Lab)
Sensitivity • Smart Monte Carlo method
FEAPpv
• Local Reliability (FORM, SORM,..)
Reliability • Global Reliability (Active learning-based)
OpenSeesPy • Importance Sampling
(general python interface)
Parameter • OPT++GaussNewton
Estimation • NL2SOL
Custom Blue: In-house developed
Bayesian program • DREAM
Calibration • TMCMC

Surrogate Custom UQ • Custom UQ algorithm

Model Surrogate • Gaussian process surrogate modeling
Modeling • Gaussian process multi-fidelity modeling
• Probabilistic learning on manifolds
Sang-ri Yi yisangri@berkeley.edu 9
 quoFEM (v.3.0)
Algorithms
FEM Engines Methods • Latin Hypercube Sampling
• Monte Carlo
Sampling • Gaussian Process Regression
OpenSees • Polynomial Chaos Expansion
Global • Probability model-based approximation
FEAPpv Sensitivity • Smart Monte Carlo method
• Local Reliability (FORM, SORM,..)
Reliability • Global Reliability (Active learning-based)
OpenSeesPy • Importance Sampling
(general python interface)
Parameter • OPT++GaussNewton
Estimation • NL2SOL
Custom Bayesian • DREAM
Calibration • TMCMC

Surrogate Custom UQ • Custom UQ algorithm

Model Surrogate • Gaussian process surrogate modeling
Modeling • Gaussian process multi-fidelity modeling
• Probabilistic learning on manifolds
Sang-ri Yi yisangri@berkeley.edu 10
 Reliability Analysis
Dakota theory manual Section 2.1
Gives CDF value (1 − 𝑃𝑓 )

▪ Local reliability
- Mean value (MVFOSM)
- Most Probable Point (Design point)
- Exact MPP search – traditional FORM/SORM
- Approximate MPP search – faster convergence, less accuracy

Optimization: sequential quadratic programming (SQP) and nonlinear interior-

point (NIP) optimization

Sang-ri Yi yisangri@berkeley.edu 11
 Reliability Analysis
Dakota theory manual Section 2.2.2, Chapter 8
▪ Global reliability
- Use surrogate model (Gaussian process) to approximate limit-state function
→ Works well when the limit-state function is multi-modal / highly nonlinear

Sang-ri Yi yisangri@berkeley.edu 12
 Reliability Analysis
Dakota theory manual Section 2.2.1
▪ Importance sampling
- Basic/adaptive/multimodal-adaptive

Sang-ri Yi yisangri@berkeley.edu 13
 Global Sensitivity Analysis
▪ Smart Monte Carlo (Dakota Engine)
Two independent N sample set
𝐴 𝒙 𝟏 𝑦 (1) 𝐵 ෥(𝟏)
𝒙 𝑦෤ (1)
𝒙 (𝟐) 𝑦 (2) ෥(𝟐)
𝒙 𝑦෤ (2)
… … … …
N
𝑦 N ෥(𝑵)
𝒙 𝑦෤
𝒙(𝑵)

Designed sample set to estimate Sobol indices of 𝑋1

Case 1 Case 2
1
𝐴𝐵1 𝑥෤11 𝑥෤2 … 𝑥෤𝐷1 𝑦ො 1 Y Designed resampling Y Designed resampling
to evaluate Si to evaluate Sj
… … … Sampling

xi xj

Sang-ri Yi yisangri@berkeley.edu 14
 Global Sensitivity Analysis
▪ Probability-mode based GSA (SimCenter UQ engine)
Estimation algorithm
• Approximate joint distribution of 𝑓 𝑋𝑖 , 𝑌 using a Gaussian
𝑁 𝑥1 𝑥2 … 𝑥𝑖 … 𝑥𝐷 𝑦 mixture model (GMM)
• Estimate 𝔼[Y|𝑋𝑖 ] from GMM 𝑓 X𝑖 , 𝑌
(𝑛)
• Repeat for different 𝑋𝑖 samples to get sample variance
(𝑛)
𝕍ar𝑥𝑖 𝔼 𝒙𝑖 ̅ Y|𝑋𝑖

→ Supports group-wise sensitivity index

Sang-ri Yi yisangri@berkeley.edu 15
 Example1
Structure (Opensees) Excitation
Rinaldi
wR 7 near-field

w 6
El Centro
w 5 far-field

w 4
Random Name Mean C.O.V
w 3
Variables w 100 0.1
wR 50 0.1
w 2 Steel 02 Material k 326 0.1
Fy 50 0.1
alpha 0.2 0.1
factor (PGA) 0.1 0.1

Sang-ri Yi yisangri@berkeley.edu 16
 Example2
Structure (Opensees)

Random Variables: Quantity of Interest:

k1,k2 Eigenvalue (rad/s) 2

Sang-ri Yi yisangri@berkeley.edu 17
 Example3
Truss model written in OpenseesPy

Sang-ri Yi yisangri@berkeley.edu 18