28th, June, 2010, 28th AIAA Applied Aerodynamics Conference Design Optimization Utilizing Gradient/Hessian Enhanced Surrogate Model Wataru YAMAZAKI, Markus P. RUMPFKEIL, Dimitri J. MAVRIPLIS Dept. of Mechanical Engineering, University of Wyoming, USA Outline *Background - Efficient CFD Gradient/Hessian calculations - Surrogate Model Enhanced by Gradient/Hessian - Uncertainty Analysis *Objectives *Surrogate Model Approaches - Kriging - Direct and Indirect Gradient-enhanced Kriging - Gradient/Hessian-enhanced Kriging Approaches *Results & Discussion - Analytical Function Fitting - Aerodynamic Data Modeling - 2D Airfoil Drag Minimization - Uncertainty Analysis at Optimal Airfoil *Conclusions -2- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Background~ Efficient CFD Hessian Calculation An efficient CFD Hessian calculation method by Adjoint method and Automatic Differentiation (AD) For steady flow i. Solutions for grid deformation / flow residual equations sD, xD 0 ii. RD, xD, wD 0 Adjoint solutions for flow / grid deformation equations T T R F 0 w w T T s F T R 0 x x x iii. Ndv linear solutions each for dx/dDj and dw/dDj iv. Ndv(Ndv+1)/2 cheap evaluations for each Hessian component 2 d F dD j dDk jk F jk R jk s T T M.P. Rumpfkeil and D.J. Mavriplis, AIAA-2010-1268 “Efficient Hessian Calculations using Automatic Differentiation and the Adjoint Method” -3- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Background~ Efficient CFD Hessian Calculation An efficient CFD Hessian calculation method by Adjoint method and Automatic Differentiation (AD) Grid Deformation Flow Residual Flow Adjoint Mesh Adjoint dx/dD1 dw/dD1 dx/dD2 dw/dD2 ...... dx/dDNdv dw/dDNdv Gradient and Hessian M.P. Rumpfkeil and D.J. Mavriplis, AIAA-2010-1268 “Efficient Hessian Calculations using Automatic Differentiation and the Adjoint Method” -4- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Background~ Approximate CFD Hessian For steady flow, a special form of objective function F K w F 2 1 k k target 2 k F k 1 e.g. F dFk wk dDi dD j k 1 dDi 2 d F K dFk wk k 1 dDi K w C 2 1 L L C target 2 L wD C D C target 2 D T 2 K dFk d Fk target dD wk Fk Fk dDi dD j k 1 j T dFk dD j 0 Last approximation is accurate only nearly optimum Approximate Hessian only requires the first-order derivatives M.P. Rumpfkeil and D.J. Mavriplis, AIAA-2010-1268 “Efficient Hessian Calculations using Automatic Differentiation and the Adjoint Method” -5- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Background~ Uncertainty Analysis Uncertainty due to manufacturing tolerances in-service wear-and-tear etc Analysis of mean/variance/PDF of objective function w.r.t. fluctuation of design variables Full Monte-Carlo Simulation Thousands/Millions exact function calls Accurate and easy, but computationally expensive Moment Method Taylor series expansion by grad/Hessian at the center No information about PDF Inexpensive Monte-Carlo Simulation Thousands/Millions surrogate model function calls Much lower computational cost -6- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Objectives The efficient adjoint gradient/Hessian calculation methods will be effective… for more efficient global design optimization with G/H-enhanced surrogate model approach for more accurate and cheaper uncertainty analysis by inexpensive Monte-Carlo simulation with G/H-enhanced surrogate model Development of gradient/Hessian-enhanced surrogate models Application to design optimization and uncertainty analysis -7- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Kriging, Gradient-enhanced Kriging Kriging model approach - originally in geological statistics Two gradient-enhanced Kriging (cokriging or GEK) Direct Cokriging Gradient information is included in the formulation (correlation between func-grad and grad-grad) Indirect Cokriging Same formulation as original Kriging Additional samples are created by using the gradient info Kriging model by both real and additional pts 2D example : Real Sample Point : Additional Sample Point -8- x add x i x yadd yxi x T yxi x Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Gradient/Hessian-enhanced Kriging Indirect Approach x add x i x yadd yxi x G T 1 x Hx T 2 2D example : Real Sample Point : Additional Sample Point Arrangements to Use Full Hessian / Diagonal Terms Major parameters : distance between real / additional pts number of additional pts per real pt Worse matrix conditioning with smaller distance larger number of additional pts Severe tradeoffs for these parameters -9- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Gradient/Hessian-enhanced Kriging Direct Approach Consider a random process model estimating a function value by a linear combination of function/gradient/Hessian components n n yˆ x i 1 wi y i i 1 n i yi i yi ' '' i 1 Minimizing Mean-Squared-Error (MSE) between exact/estimated function n MSE yˆ x E w i y i i 1 with an unbiasedness constraint n i 1 n i y i y Y i 1 ' i '' i 2 n w i 1 i 1 Solving by using the Lagrange multiplier approach J J J J n J MSE yˆ x w i 1 0 wi i i i 1 -10- for i Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Gradient/Hessian-enhanced Kriging Direct Approach Introducing correlation function for covariance terms Correlation is estimated by distance between two pts with radial basis function Cov Z xi R x , x Z x j R x i , x j 2 Cov Z x i , , Z x j x k 2 i j x j k Unknown parameters are determined by the following system of equations R T F F x r ~ 0 1 x T w1 , , 1 , , 1 , Final form of the gradient/Hessian-enhanced direct Kriging approach is T 1 yˆ x r x R Y F F R F T 1 1 F T 1 R Y r x R mean constant term correlatio ns between x and observed Y y1 , , y1 , , y1 , ' '' T correlatio ns between observed data data exact informatio n at given samples -11- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Gradient/Hessian-enhanced Kriging Direct Approach R x1 ,x1 R x n ,x1 R x ,x 1 1 1 x1 R R x n , x 1 N dv x n 2 R x1 ,x1 2 1 x1 2 R x n , x 1 2 Ndv x n R x1 ,x n R x n ,x n T 1 yˆ x r x R Y F R x1 ,x1 x1 R x1 ,x n 1 R x n , x 1 R x1 ,x1 x n dv R x n , x n x1 2 R x1 ,x n N 1 2 R x n , x 1 2 2 x n 2 R x n , x n Ndv 2 n n n n n n x Correlations between F-F, F-G, G-G, F-H, G-H and H-H R R R R R function Up to 4th xorderderivatives x x x x x x xof correlation by TAPENADE Differentiation Automatic R R R R R No sensitive parameter x x x x x x x x R R thanindirect R matrix R R conditioning Better approach x1 ,x n x1 1 2 x n , x n 2 2 x1 ,x n x 2 R x n , x n 2 1 1 x n 2 Ndv N dv n 3 1 1 2 1 1 1 1 x n d v x1 N 1 N dv n 4 x 1 , x n x n d v x n d v x1 x n d v 2 N -12- N dv n N 2 2 1 1 2 1 2 N N dv n N dv n 4 x 1 1 1 1 x n , x n 2 x1 ,x1 x x 4 R x 1 , x n 1 1 3 N dv n 1 1 x n , x 1 N dv n 2 x 1 , x n x x 3 R x n , x n 2 3 2 3 1 1 1 1 Ndv n 2 x1 ,x1 2 x n , x n 3 x1 ,x1 3 N dv n 2 1 1 1 2 x 1 , x n 1 1 x x 3 R x n , x 1 2 2 2 x n , x 1 N dv n 2 N x1 ,x1 1 1 x1 x n dv 2 x n , x 1 N dv n x1 2 1 R x n , x n 4 x n d v x n d v 2 N 2 N Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Infill Sampling Criteria for Optimization How to find promising location on surrogate model ? Maximization of Expected Improvement (EI) value Potential of being smaller than current minimum (optimal) Consider both estimated function and uncertainty (RMSE) y min ˆy x y min ˆy x ˆ y x s s s Exact Function EI 0, ˆ y Sample Points Kriging 0 s EI RMSE EI 1.0 1.0E-02 0.8 8.0E-03 0.6 6.0E-03 0.4 4.0E-03 0.2 2.0E-03 0.0 0.0E+00 EI Function / RMSE EI x y min 0.0 0.2 -13- 0.4 0.6 Design Variable 0.8 1.0 Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Results & Discussion 2D Rastrigin Function Fitting y x 20 x1 x 2 10 cos 2 x1 cos 2 x 2 2 2 80 samples by Latin Hypercube Sampling Direct Kriging approach Exact Rastrigin Function Gradient/Hessian-enhanced Function-based Gradient-enhanced Kriging -15- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 5D Rosenbrock Function Fitting F: FG: FGHd: FGH: Function-based Kriging Gradient-enhanced G/diag. Hess-enhanced G/full Hess-enhanced RMSE .vs. Number of sample points Superiority in direct Kriging approaches thanks to exact enforcement of derivative information better conditioning of correlation matrix -16- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Validation on Rosenbrock Func. 1 x N d v 1 y x i 2 2 2 100 x i 1 x i i 1 CDFs of Full-MC and IMC 1.0 1.E+03 0.9 1.E+02 0.8 F 1.E+01 CDF Objective Function Optimization History 1.E+04 Direct_FG 1.E+00 Direct_FGH Indirect_FGH 1.E-01 Full-MC 0.7 IMC_F 0.6 IMC_FG IMC_FGH 0.5 1.E-02 0.4 1.E-03 0.3 0 100 200 300 0 Number of Sample Points Minimization of 20D Rosenbrock 30 initial sample points by LHS EI-based infill sampling criteria Faster convergence in G/H-enhanced direct approach 10 20 30 40 50 60 70 80 90 Function Value Uncertainty analysis on 2D Rosenbrock 5 sample points for surrogate model (No sample point on the center location) Superior performance in G/H-enhanced Inexpensive MC (IMC) -17- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Aerodynamic Data Modeling Unstructured mesh CFD Steady inviscid flow, NACA0012 20,000 triangle elements Mach Number [0.5, 1.5] Angle of Attack[deg] [0.0, 5.0] 21x21=441 validation data Exact Hypersurface of Lift Coefficient -18- Exact Hypersurface of Drag Coefficient Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Aerodynamic Data Modeling Exact Function-based Gradient-enhanced Kriging Cl Cd Adjoint gradient is helpful to construct accurate surrogate model CFD Hessian is not helpful due to noisy design space -19- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 2D Airfoil Shape Optimization Unstructured mesh CFD Steady inviscid flow, M=0.755 NACA0012, 16 DVs for Hicks-Henne function Objective function of inverse design form F w C C 2 1 l l target 2 l wd Cd C 1.0 target 2 d 0.9 0.8 0.7 1 2 Cl 0.675 2 100 2 Cd 0.000 2 H(x) 0.6 0.5 0.4 0.3 0.2 0.1 0.0 Exact / Approximate CFD Hessian available Computational time of F : 2 min, FG : 4 min, FGHapprox. : 6 min, FGHexact : 36 min (4 min in parallel) Geometrical constraint for sectional area 0.0 -20- 0.1 0.2 0.3 0.4 0.5 x 0.6 0.7 0.8 0.9 Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 1.0 2D Airfoil Shape Optimization Start from 16 initial sample points which only have function info Gradient/Hessian evaluations only for new optimal designs Faster convergence in derivative-enhanced surrogate model Best design in gradient/exact Hessian-enhanced model -21- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 2D Airfoil Shape Optimization NACA0012 (Baseline) Optimal by G/exact H-enhanced model Towards supercritical airfoils Shock reduction on upper surface -22- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 2D Airfoil Shape Uncertainty Analysis Geometrical uncertainty analysis at optimal airfoil Center = optimal obtained by Grad/exact H model optimal (center) ±0.1 airfoil Comparison between 2nd order Moment Method (MM2) using gradient/Hessian at the center Inexpensive Monte-Carlo (IMC1) using final surrogate model obtained in optimization Inexpensive Monte-Carlo (IMC2) using different G/H-enhanced model by 11 samples Full Non-Linear Monte-Carlo (NLMC) using 3,000 CFD function calls -23- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 2D Airfoil Shape Uncertainty Analysis MM2 using derivative at the center IMC1 using G/H surrogate model obtained in optimization IMC2 using different G/H model by 11 samples (for st.devi.=0.01) NLMC using 3,000 CFD function calls Mean of objective w.r.t. standard deviation of all design variables IMC showed good agreement with NLMC at smaller st. devi. Necessity of additional sampling criteria for total model accuracy ? Promising IMC with much cheaper computational cost -24- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Concluding Remarks / Future Works Development of gradient/Hessian-enhanced Kriging models Application to design optimization and uncertainty analysis Direct Kriging approach is superior to indirect approach More accurate fitting on exact function Faster convergence towards global optimal design Promising inexpensive Monte-Carlo simulation at much lower cost Application to higher-dimensional / complicated design problem Robust design with inexpensive Monte-Carlo simulation Gradient/Hessian vector product-enhanced approach Thank you for your attention !! -25- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Appendix Moment Method Taylor series expansion by grad/Hessian at the center No information about PDF 1st order Moment Method MM 1 F x c MM 1 2 N dv i 1 dF dD i D i xc 2 2nd order Moment Method MM 2 MM 1 1 N dv 2 i 1 2 MM 2 2 MM 1 1 2 Di xc 2 d F dD dD j 1 i N dv N dv 2 i 1 -27- d 2F dD i2 DD i j xc j 2 Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Gradient/Hessian-enhanced Kriging Implementation Details Correlation function of a RBF 1 6 2 2 1 h 35 h 18 h 3 scf , h 3 0 for h 1 else Estimation of hyper parameters by maximizing likelihood function with GA Correlation matrix inversion by Cholesky decomposition Search of new sample point location by maximizing Expected Improvement (EI) value with GA -28- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Infill Sampling Criteria for Optimization How to find promising location on surrogate model ? Expected Improvement (EI) value Potential of being smaller than current minimum (optimal) Consider both estimated function and uncertainty (RMSE) y min ˆy x y min ˆy x ˆ y x s s s Exact Function EI 0, ˆ y Sample Points Kriging 0 s EI RMSE EI 1.0 1.0E-02 0.8 8.0E-03 EI-based criteria have good balance between global/local searching 0.6 6.0E-03 EI Function / RMSE EI x y min 0.4 4.0E-03 0.2 2.0E-03 0.0 0.0E+00 0.0 0.2 -29- 0.4 0.6 Design Variable 0.8 1.0 Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 5D Rosenbrock Function Fitting # of pieces of information = sum of # of F/G/H net components To scatter samples is better than concentration at limited samples Approximated computational time factor N sample TF T , i Ti 1 / 2 / 3, if i has F / FG / FGH i 1 G/H-enhanced surrogate model provides better performance with efficient Gradient/Hessian calculation methods -30- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 1D Step Function Fitting 1.5 Function Value 1.0 Exact Samples 0.5 F FG FGH 0.0 -0.5 0.0 0.2 0.4 0.6 0.8 1.0 Design Variable Much better fit by G/H-enhanced direct Kriging -31- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Minimization of 20D Rosenbrock Func. 1.E+04 Objective Function 1.E+03 F FG FGH 1.E+02 1.E+01 1.E+00 1.E-01 1.E-02 0 3000 6000 9000 12000 15000 Computational Time [sec] Minimization of 20 dimensional Rosenbrock function No computational cost for Func/Grad/Hess evaluation Expensive for construction - likelihood function maximization - inversion of correlation matrix Parallel computation for the likelihood maximization problem -32- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Uncertainty Analysis 150 1.0 130 0.9 Full-MC 110 90 MM2 70 IMC_F 50 IMC_FG 0.6 IMC_FGH 0.5 30 CDF at St. Devi.=0.15 0.8 CDF Mean of Function Uncertainty analysis at (1.0,1.0) on 2D Rosenbrock 5 sample points for surrogate model approaches (No sample point on the center location) 2nd order Moment Method (MM2) by G/H at the center Superior results in G/H-enhanced Inexpensive MC (IMC) 0.7 0.4 10 -10 0.3 0.0 0.1 0.2 0.3 0.4 0.5 0 Standard Deviation of DVs 10 20 30 40 50 60 70 80 90 Function Value St. Devi. = 0.15 means the possibility within -0.15<dx<0.15 is about 70% -33- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming Aerodynamic Data Modeling NACA0012 M=1.4 AoA=3.5[deg] Noisy in Mach number direction 0.212 0.1090 CFD Data CFD Data 0.1088 Linear by Adj_Grad 0.211 Quadratic by Adj_G/H 0.1086 CD CL Quadratic by Adj_G/H Linear by Adj_Grad 0.210 0.1084 0.209 0.208 1.390 0.1082 1.395 1.400 Mach Number 1.405 1.410 Cl 0.1080 1.390 1.395 1.400 Mach Number 1.405 1.410 Cd -34- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming 2D Airfoil Shape Uncertainty Analysis Cumulative Density Function at St. Devi. of 0.01 Quadratic model only by using gradient/Hessian at optimal Additional sampling criteria to increase total model accuracy -35- Yamazaki, W.,Wataru Dept. YAMAZAKI, of Aero. Eng., Tohoku Univ. Univ. of Wyoming