SEPARABLE SAMPLING OF THE LIMIT STATE FOR ACCURATE MONTE CARLO SIMULATION Bharani Ravishankar, Benjamin Smarslok Advisors Dr. Raphael T. Haftka, Dr. Bhavani V. Sankar Motivation - Probability of Failure Problems Monte Carlo simulation-based techniques can require expensive calculations to obtain random samples Capacity R C To improve the accuracy of pf estimate for complex limit states without performing additional expensive response computation? 2 Outline & Objectives Review Monte Carlo simulation techniques - Crude Monte Carlo method - Separable Monte Carlo method Simple limit state example - Explain the advantage of regrouping random variables Complex (non-separable) limit state example - Tsai Wu Criterion -Demonstrate regrouping & separable sampling of stress and strength Compare the accuracy of the Monte Carlo methods Conclusions 3 Monte Carlo Simulations Common way to propagate uncertainty from input to output & calculate probability of failure Limit state function is defined as R C, Failure R( X1 ) C ( X2 ) R C, Safe Response depends on a set of random variables X1 Capacity depends on a set of random variables X2 R Response (eg. Stress) C Capacity (eg.Yield Strength) R Crude Monte Carlo (CMC) - most commonly used pˆ cmc 1 N N I R C i C Potential failure region i i 1 4 Crude Monte Carlo Method Assuming Response ( ) involves Expensive computation (FEA) Limit state function • isotropic material • diameter d, thickness t • Pressure P= 100 kPa max Y 0 max 2d P t Failure max Y z y Random variables Response - Stress = f (P, d, t) Capacity - Yield Strength, Y x hoop 100 kPa R axial C Y CV pˆ cmc 0.21 Example: Pf estimate 0.19 Y C : N 13, 1.5 N 10 p f 0.062 1 pf pf N 0.17 0.15 0.13 0.11 0.09 0.07 0 I – Indicator function takes value 0 (not failed) or 1( failed) 2000 4000 6000 Number of samples 5 8000 Separable Monte Carlo Method If response and capacity are independent, we can use all of the possible combinations of random samples Empirical CDF pˆ smc 1 MN N M I R C i pˆ smc j i 1 j 1 1 N N Fˆ (R ) C i 1 Example: C Y N 10 M 10 p f 0.062 0.21 CMC 0.19 Pf estimate R 0.17 SMC 0.15 0.13 0.11 0.09 0.07 0 2000 4000 6000 Number of samples 6 8000 i Regrouping the random variables Random variables Response - Stress = f (P, d, t) Capacity - Yield Strength, Y max 2d P t Regrouping the random variables Stress is a linear function of load P P, d, t and Y are independent random variables uP u – Stress per unit load Regrouped variables Stresses per unit load u Pressure load P Yield Strength Y 7 Monte Carlo Simulation Summary Crude MC traditional method for estimating pf – Looks at one-to-one evaluations of limit state – Expensive for small pf Separable MC uses the same amount of information as CMC, but is inherently more accurate – Use when limit state components are independent – Looks at all possible combinations of limit state R.V.s – Permits different sample sizes for response and capacity For a complex limit state, the accuracy of the pf estimate could be improved by regrouping and separable sampling of the RVs 8 Complex limit state problem Determination of Stresses z y Material Properties E1,E2,v12,G12 x Ny 100 kPa Laminate Stiffness (FEA) Nx Pressure vessel -1m dia. (deterministic) Thickness of each lamina 0.125 mm (deterministic) Lay up- [(+25/-25)]s Internal Pressure Load, P= 100 kPa Strains x , y xy y x Stress x , y Stress in each ply 1 , 2 ,12 9 Loads P Limit State - Tsai-Wu Failure Criterion Non-separable limit state G( , S) F1112 F22 22 F66122 F11 F2 2 2F121 2 1 No distinct response and capacity Random Variables F = f (Strengths S) =f (Laminate Stiffness aij, Pressure P) F11 1 1 1 F 1 S L S L S L S L F22 1 1 1 F 2 ST ST ST ST F66 F11.F22 1 F 12 2 S LT 2 obtained from Classical Laminate Theory (CLT) 1 a11 2 a12 0 12 a12 a22 0 0 P / 2 0 P / 4 aij P a66 0 F – Strength Coefficients Limit state G = f (F, ); G < 0 safe G ≥ 0 failed S – Strengths in Tension and Compression in the fiber and transverse direction 10 Estimation of probability of failure G( , S) F1112 F22 22 F66122 F11 F2 2 2F121 2 1 RVs - Uncertainty { } = {1, 2,12}T Parameters Mean E1 (GPa) 159.1 E2 (GPa) 8.3 G12 (GPa) 3.3 12 (no unit) CV% Crude Monte Carlo 5 1 N pˆ cmc I [G( i , Si ) 0] N i 1 0.253 Pressure P (kPa) 100 S1T (MPa) 2312 S1C (MPa) 1809 S2T (MPa) 39.2 S2C (MPa) 97.2 S12 (MPa) 33.2 S = {S1T S1C S2T S2C S12 } N 15 10 CV(Pressure) > CV(Strengths) > CV(Stiffness Prop.) All the properties are assumed to have a normal distribution 11 N Separable Monte Carlo pˆ smc 1 MN N M I G( , S ) 0 i i 1 j 1 N M j CMC and SMC Comparison G( , S) F1112 F22 22 F66122 F11 F2 2 2F121 2 1 Coefficient of variation 45 40 SMC 35 CMC N=500, repetitions = 10000 30 25 20 Actual Pf = 0.012 15 10 5 0 500 5000 50000 M Samples Expensive Response limited to N=500 (CLT) Cheap Capacity varied M= 500, 5000 samples 12 Tsai – Wu Limit State Function Original limit state G ( , S) pˆ smc 1 MN N M I G( , S ) 0 Stresses i i 1 j 1 Strengths S Stresses per Load P u Element Analysis unit load Finite From Statistical Finite Element AnalysisExpensive Expensive N j From Statistical distribution distribution Cheap Cheap Regrouping the expensive and inexpensive variables Regrouped limit state G( u , P,S) u pˆ sm c 1 M .N Expensive Stresses per unit load N M I G( u i , Pj , S j ) i 1 j 1 Cheap u Pressure Load P u – Material Properties, P – Pressure Loads, S – Strengths 13 Strengths S 0 M Regrouping the random variables Stresses Cost Uncertainty Material Properties Load P Strengths S Expensive Cheap Cheap ~ 5% 15% 10% G ( , S) G( u , P,S) G ( , S) Gu ( u , P,S) G( , S) Gu ( u , P,S) , u , P,S -Mean values 14 Comparison of the Methods 2 G( , S) F1112 F22 22 F6612 F11 F2 2 2F121 2 1 Expensive RVs limited to N=500 (CLT) Cheap RVs varied M= 500-50000 samples SMC M SMC-unit load CMC 45 Coefficient of Variation 40 500 1000 5000 10000 50000 Crude Monte Carlo Separable Monte Carlo CV pˆ cmc CV pˆ smc 40.0% 20.6% 18.4% 16.2% 16.0% 15.6% Separable Monte Carlo regrouped RVs 35 30 25 N=500 repetitions = 10000 20 15 10 Actual Pf = 0.012 5 0 500 5000 50000 M Samples 15 u CV pˆ smc 36.3% 26.0% 11.7% 8.2% 4.0% Accuracy of probability of failure For CMC, accuracy of pf 0.21 CMC 0.19 pf N For SMC, Bootstrapping – resampling with replacement Pf estimate CV pˆ cmc 1 pf 0.17 SMC 0.15 0.13 0.11 0.09 Initial Sample size N 0.07 0 k=1 Re-sampling with replacement, N ….…... ‘b’ bootstrap samples……….. ˆ boot pf estimate from bootstrap sample, p 4000 6000 8000 Number of samples k= b k=2 2000 Re-sampling with replacement, N ˆ boot pf estimate from bootstrap sample, p ‘b’ estimates of pˆ boot pˆ boot bootstrapped standard deviation/ CV stdev pˆ boot / CV pˆ boot stdev pˆ boot / CV pˆ boot = error in pf estimate mean pˆ boot &stdev pˆ boot Summary & Conclusions Separable Monte Carlo was extended to non-separable limit state - Tsai-Wu failure criterion. In Tsai-Wu Limit State, uncertainty in load affects the expensive stresses. By calculating response to unit loads, we can sample the effect of random loads more cheaply. Statistical independence of the random variables enables appropriate sampling, thereby improving the accuracy of the estimate. Shift uncertainty away from the expensive component furthers helps in accuracy gains. Accuracy of the methods - for the same computational cost, CMC CV% 40% SMC -original limit state SMC- Regrouped limit state 16% 4% 17