Numerical Comparison of PDE Constrained Optimization Schemes for Solving Earthquake Modeling Problems L. Velazquez¹, C. Burstede ², A.A. Sosa ¹, M. Argaez¹, O. Ghattas ² ¹ Computational Science Program, The University of Texas at El Paso (UTEP) Summary Logarithmic Barrier Method We are implementing a Primal-Dual Interior-Point method to solve an inverse seismic wave problem. We compare the numerical behavior of the logarithmic barrier (LB), primal-dual interior-point (IPM), projected conjugate gradient (PRCG), and primal-dual active sets (PDAS) methods for solving one-dimensional partial differential equations (PDE) constrained optimization problems arising on earthquake applications. Our main interest is to identify an effective strategy that allows to incorporate efficiently physical bound constraints into the problem. We run several test cases and identify that the strategy IPM works best. The problem that we present here needs from the expertise of researchers from different areas (geologists, engineers, computational scientists) to develop the capability for estimating the geological structure and mechanical properties of the earth structure. Our contribution comes to help in the development of a new optimization approach to the new inversion methods based on PDE’s. Forward and Inverse Modeling At the heart of the site characterization problem using active (or passive) dynamic sources lies an inverse wave-based problem. Inverse problems are notoriously more difficult than the corresponding forward wave propagation problems. In the forward problem, one wishes to determine the soil response due to a prescribed excitation under the assumption that the source and the material properties are known. By contrast, in the inverse problem one wants to estimate the spatial distribution of the soil properties that results in a predicted response that most closely matches the observed records generated by either an active source or a seismic signal. The forward problems are, thus, but a mere “inner iteration” of the inverse problem, which entails multiple forward wave propagation simulations. It is thus important that forward wave simulations are carried out fast and accurately. f f ( x, t ) u u ( x, t ), u u R , and force u I f u I L u on (0,T ] A( )u f AB Then the space-time discretized inverse PDE constrained problem is: min J ( , u ) s.t. A( )u f , a b (2) 1 * T * J ( , u ) (u u ) W (u u ) R( ), 2 2 and a, b are the physical bounds. The regularization term min J ( , u ) R( )varies depending on the method chosen. We R( ) || ||2 R( ) 2 log(pi ( )) s.t. A( )u f ( 3) or Total Variation ~ v (u , y, z1 , z 2 ) : log( p ( )) Au f T i y i 1 Where y is the Lagrange multiplier associated to the equality constraints. Primal-Dual Interior-Point Method T T y p1 z1 - p2 z 2 Where z1, z2 0 is the Lagrange multiplier associated with the inequality constraints. The methodology is to keep the iterates positive except at the solution of the problem. Optimality Conditions 0 We apply Newton’s method to (4) or (5), and reduce the system to obtain: 2 J C E Ty W Ey Eu A J B * A u Ay W (u u )(6) 0 y Au f T Eu T J ( , u) 2 T Eu 2 y R2 , R2 1 1 T LB : P1 e P2 e B ; T 1 1 IPM : P e P e 1 2 2 (7) 0 R( ) , and E y Ay. LB : P2 P2 1 2 C 1 1 IPM : P1 Z1 P2 Z 2 PRCG PDAS IPMTV IPM Step 2 For k=0,1,2,… until k 205 150 150 ~ k 0 LB PRCG PDAS IPMTV IPM Wild Experiment 1000 Jbest* || µtrue--µ*| LB 2.5e-5 1e-1 PRCG 1.8e-5 9.5e-2 PDAS 2.2e-5 1.1e-1 IPMTV 4.2e-6 7.5e-2 IPM 7e-7 7.7e-2 CG iterations 800 902 600 739 534 400 200 360 373 0 LB PRCG PDAS IPMTV v~ 1200 according to the method selected nonnegative integer such that k satisfies 4 J ( k k k , uk ) J (vk ) 10 k J ( 1073 800 600 Jbest* || µtrue--µ*|| LB* 2.1e-4 7e-2 PRCG ----PDAS 6.95e-5 5.7e-2 IPM 3.68e-6 2.5e-2 IPM 5.65e-7 4.4e-2 566 400 k , uk ) k 446 200 479 0 vk 1 vk k vk We use four random initial points for solving two test problems (step, wild) with two different choices of level and grid coarse (6,3) and (7,5), respectively. The numerical results of PRCG and PDAS, and the Matlab code were provided by the UT team. The experimentation shows that Interior-Point Methods, without regularization, to be competitive in terms of robustness, accuracy , and total number of conjugate gradient iterations. CG iterations 1000 t IPM Step Experiment 1 ~ (Sufficient decrease) Find k k where t is the smallest 2 179 50 according to the method selected. and calculate 196 100 J ( k , uk ) Step 5 (Force positivity) Calculate 251 according to the method selected Step 4 Solve (7) for step Step 6 CG Iterations 200 Step 1 Select a method j=LB or IPM. Choose an initial point 0 Jbest* || µtrue--µ*|| LB 8.60e-5 1.5e-2 PRCG 1.82e-5 1e-4 PDAS 1.82e-5 1e-4 IPMTV 2.25e-6 3.1e-4 IPM 1.49e-6 1.2e-5 250 Algorithm and IPM Step Experiment does not change when either method is implemented T ~ g Eu y R1 B v0 ( 0 , u0 , y0 , z10 , z 20 ) 188 50 300 Numerical Results 2 ~ g Jbest* ||µtrue-µ*|| LB 1.82e-5 2.5e-3 PRCG 1.79e-5 4e-4 PDAS 1.79e-5 4e-4 IPMTV 2.37e-6 3.8e-5 IPM 4.27e-12 4.7e-6 232 159 LB Step 8 Return to Step 2 235 100 2. The only differences between the systems (6) and (7) are the terms C and B that depend only on the choice of the method LB or IPM. Step 7 (Update) Set 1 1 z1 z1 P1 e Z1 , z 2 z 2 P2 e Z 2 IPM H g ~ T T 1 H R2 Eu A WA Eu Step 3 Choose T Eu y R1, Eu 302 150 or we can apply Gauss-Newton’s method having: 3. z1 z 2 J ( , u ) T IPM 1 1 J ( , u ) P e P e A( ) y W (u * u ) 1 2 0, (5) * F ( , u, y ) A( ) y W (u u ) 0 , ( 4 ) F ( , u , y , z , z ) 1 2 A ( ) f IPM P Z e A ( ) f 1 1 IPM P2 Z 2 e For 300 200 Remarks: ~ 1. When Newton’s method is implemented H has second order information while Gauss-Newton use only first order information. The lagrangian associated to (2) is: CG Iterations ~ 2 T T 1 T T T T 1 H Eu y R2 E y A Eu Eu A E y Eu A WA Eu 2n ( , u , y, z1 , z 2 ) J ( , u ) Au f and then solve for We can solve (7) using Newton’s method where: The lagrangian function associated to (3) is T 350 ˜ C and g g ˜B where H H where p(μ ) [ p1,p2 ] T [μ a,b μ ] T . ( , u , y ) J ( , u ) Wild Experiment 250 IPM Nonlinear System (1) We can reduce the nonlinear system (6) in terms of i 1 Where can select Tikhonov regularization regularization 2n T u 0 on {t 0}, u 0 on {t 0} Where J ( , u) A( )u, R1 R( ), Z1 diag(z1), Z 2 diag(z 2 ) , P1 diag(p1), P2 diag(p2 ), and 2 e = (1,,1)T . subject to additional boundary conditions: u u u T shear-modulus We solve problem (2) through a sequence of log-barrier sub-problems as parameter goes to zero. Where F is a square nonlinear function, The forward problem is based on the elastic wave equation for n Reduced Nonlinear System ( p1 , p2 , z1 , z 2 ) 0, Seismic Wave Propagation Problem displacement (state) ²Center for Computational Geosciences and Optimization, The University of Texas at Austin (UTA) LB References PRCG PDAS IPMTV IPM 1. Numerical Comparison of Constrained PDEs Optimization Schemes for Solving Earthquake Modeling Problems. L. Velazquez, C. Burstede, A.A. Sosa, M. Argaez, and O. Ghattas, Technical Report. 2. Numerical Optimization. J. Nocedal, and S. J. Wright, . 3. An Introduction to Seismology Earthquakes and Earth Structure, S. Stein and M. Wysession, Blackwell Publishing, , 2006 Contact Information Acknowledgements Uram A. Sosa, Graduate Student This work is being funded by NSF Grants Computational Science Program NEESR-SG CMMI-0619078 and Crest usosaaguirre@miners.utep.edu Cyber-ShARE HRD-0734825.