Goal-oriented adaptive finite element methods for quasilinear PDE Numerical comparison of methods Sara Pollock Collaborators: Michael Holst, Yunrong Zhu Texas A&M Department of Mathematics October 24, 2014 S. Pollock Texas A&M Mathematics 1/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Goal-oriented method Goal function g : X → R, find g (u ) where u ∈ X satisfies the primal problem: F (u ) = −div(κ(u )∇u ) − f = 0. Assume a unique solution. Following [Becker, Rannacher, 2001]. Trivial constrained optimization problem for g: Find u ∈ X such that g (u ) = min g (v ). {v ∈X :F (v )=0} The critical points of the Lagrangian correspond to the minimum u, L (u , z ) := g (u ) − hF (u ), z iY ∗ ,Y , where z is the adjoint variable. Euler-Lagrange system: hF (u ), v iY ∗ ,Y = 0, 0 for all v ∈ Y 0 hF (u )w , z iY ∗ ,Y = hg (u ), w iX ∗ ,X , for all w ∈ X . If g is a linear functional, the right hand side is hg , w iX ∗ ,X . Discrete Euler-Lagrange system: hF (uh ), v iY ∗ ,Y = 0, 0 for all v ∈ Yh 0 hF (uh )w , zh iY ∗ ,Y = hg (uh ), w iX ∗ ,X , S. Pollock Texas A&M Mathematics for all w ∈ Xh . 2/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Error representation and convergence Denote the weak-form residual corresponding to the primal and approximate dual problems ρ(uh , v ) := −h(F (uh ), v i, for all v ∈ Y ρ (zh , w ) := hg 0 (uh ), w i − hF 0 (uh )w , zh i, for all w ∈ X . ∗ where the Gateaux derivative in direction w given as F 0 (u )w = ddε F (u + εw ) Then ε=0 as in [Becker, Rannacher, 2001] we have the error representation g (u ) − g (uh ) = 1 2 1 ρ(uh , z ) + ρ∗ (zh , u ) + R3 , 2 where R3 is given by Z 1 R3 = L 000 (uh + se; zh + se∗ )({e, e∗ })({e, e∗ })({e, e∗ })s(s − 1)ds 0 Z = 1 hg 000 (uh + se)(e)(e), ei − hF 000 (uh + se)(e)(e)(e), zh + se∗ i −3hF 00 (uh + se)(e)(e), e∗ i s(s − 1) ds. 0 S. Pollock Texas A&M Mathematics 3/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Results for adaptive methods. Example: stationary heat equation with nonlinear diffusion: F (u ) = −div(κ(u )∇u ) − f = 0 in Ω, u = 0 on ∂Ω For simplicity say Vh ⊂ X = Y . Discrete system: hκ(uh )∇uh , ∇v i = hf , v i, hκ(uh )∇w , ∇zh i + hκ0 (uh )w ∇uh , ∇zh i = hg , w i, for all v ∈ Vh for all w ∈ Vh . If κ(s) ∈ C 2 (R) satisfies κ(s) ≥ α ≥ 0 and κ(s) along with its first and second derivatives are bounded, the adaptive algorithm for u converges. Weak convergence of the primal residual implies convergence of the weak solution [Holst, Tsogtgerel, and Zhu, 2009]. Z ρ(uh , v ) := − F (uh )v dx → 0, for all v ∈ Y =⇒ uh → u . Claim: A similar idea can be used to show convergence of the Euler-Lagrange system for the goal-oriented method. Question: Does it matter?? Convergence of uh → u implies g (uh ) → g (u ), at least eventually. S. Pollock Texas A&M Mathematics 4/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Numerical comparison of methods Investigate numerically the effectiveness of the goal-oriented method for quasilinear problems. We will compare numerical results using three different adaptive algorithms 1 PR + DR: Refine for the local strong-form residuals of both the primal and dual problems. 2 DWR: Refine for the product of primal residual and dual solution, discussed in, e.g., [Becker, Rannacher, 2001] (among many others). 3 PR : Refine for the residual of the primal problem (standard AFEM). 4 PR + DWR : DWR refinement plus refinement for the primal residual. The GOAFEM algorithm is based on the standard AFEM iterative loop: SOLVE → ESTIMATE → MARK → REFINE . S. Pollock Texas A&M Mathematics 5/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Estimate Residual-based indicators: linear finite elements for both primal and dual problems. DWR: quadratic finite elements for the dual problem. VT : linear Lagrange finite elements P1 (T ) V2T : quadratic Lagrange finite elements P2 (T ) Estimate: The DWR indicator estimates the influence of the dual solution on the primal residual. Elementwise 1 ηTk (v , T ) := hR (v ), z 2 − Ik z 2 iT + hJT (v ), z 2 − Ik z 2 i∂T , 2 v ∈ VTk Residual error indicator: ηpT (v , T ) := hTp kR (v )kpL2 (T ) + hT kJT (v )kpL2 (∂T ) , v ∈ VT , with R (v ) = −(∇ · (κ(v )∇v ) + f ). The dual residuals are defined analogously. The jump residual: JT (v ) := J[κ(v )∇v ] · nK∂T and JφK∂T := lim φ(x + tn) − φ(x − tn) t →0 where n is the the outward normal defined piecewise on ∂T . S. Pollock Texas A&M Mathematics 6/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Test Problem Quasilinear diffusion problem on Ω = (0, 1)2 . Z κ(u )∇u · ∇v dx = fv dx , κ(u ) = 1 + Ω Ω The goal function g (u ) = R 1 Z Ω gu dx, 10−2 + (u − a)2 with g (x , y ) = 200 exp −200 (x − 0.7)2 + (y − 0.7)2 and f (v ) chosen so the exact solution u (x , y ) = sin(πx ) sin(πy ). The parameter a controls the smoothness and positioning of the near-singularity of the problem data in the domain. a = 0.3 (P1) a = 0.8 (P2) The initial mesh has 288 elements. The Dörfler parameter θ = 0.5. S. Pollock Texas A&M Mathematics 7/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Convergence problem!! Z κ(u )∇u · ∇v = Z fv , κ(u ) = 1 + 1 10−2 + (u − a)2 The standard Newton iterations for solving the nonlinear problem: hF (U ), vh i = 0 : Starting with initial guess U = U0 , iterate: Find: w such that hDF (U )w , v i = −hF (U ), v i Update: U = U + w Check: if kF (U )k < tolerance Sensitivity to initial data: This method did not converge for a ∈ (0.155, 1.35). In fact , starting the Newton method with the solution U0 = sin(πx ) sin(πy ) converged in 3 iterations, but failed to converge with a perturbation of more than 2 × 10−2 . For the results here a damped Newton method is implemented where the step size is decreased by a factor of 1/2 until the residual decreases (line search method). A better method by means of Tikhonov regularization with appropriate penalty parameters is currently under investigation. S. Pollock Texas A&M Mathematics 8/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Numerics: P1- error reduction Z κ(u )∇u · ∇v = 1 Z fv , κ(u ) = 1 + −2 10 + (u − 0.3)2 g (x , y ) = 200 exp −200 (x − 0.7)2 + (y − 0.7)2 Goal Error Reduction f −1 10 6000 4000 −2 10 2000 0 −2000 −3 −4000 10 −6000 1 0.8 1 0.6 0.8 0.6 0.4 −4 10 0.4 0.2 2 10 3 10 4 10 5 0.2 0 10 0 Figure: Left: goal error after 12 PR+DR (blue), 11 DWR (green), 12 PR (magenta), 10 PR+DWR (cyan) iterations, compared with n−1 . Right: problem data f (x , y ), 0 ≤ x , y ≤ 1. S. Pollock Texas A&M Mathematics 9/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Numerics: P1- mesh Z κ(u )∇u · ∇v = 1 Z κ(u ) = 1 + −2 10 + (u − 0.3)2 g (x , y ) = 200 exp −200 (x − 0.7)2 + (y − 0.7)2 Finite element mesh fv , Finite element mesh Finite element mesh Figure: Mesh after 6 adaptive refinements. Left: PR+DR; center: DWR; right: PR + DWR S. Pollock Texas A&M Mathematics 10/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Numerics: P1- global error The maximum value of uh (x , y ) should approach 1. ITER PR + DR DWR PR PR + DWR 2 4 6 8 10 0.948413 0.991561 1.00093 1.00046 0.999875 0.945183 0.965888 0.979592 0.98854 0.991947 0.947803 0.990919 1.00099 1.00037 0.999748 0.954402 0.985238 0.993353 0.997274 0.998994 An accurate height of the sinusoid in this case agrees with an accurate approximation of the quantity of interest. Some global resolution may be necessary to approximate the quantity of interest: in this case the primal solution is smooth and has global support on the domain. S. Pollock Texas A&M Mathematics 11/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Numerics: P2- error reduction Z κ(u )∇u · ∇v = 1 Z fv , κ(u ) = 1 + −2 10 + (u − 0.8)2 g (x , y ) = 200 exp −200 (x − 0.7)2 + (y − 0.7)2 Goal Error Reduction f −2 10 4000 3000 2000 1000 −3 10 0 −1000 −2000 1 0.8 1 0.6 0.8 0.6 0.4 −4 0.4 0.2 10 2 10 3 10 4 10 0.2 0 5 0 10 Figure: Left: Goal error after 13 PR+DR (blue), 16 DWR (green), 12 PR (magenta), 12 PR+DWR (cyan) iterations, compared with n−1 . Right: data f (x , y ), 0 ≤ x , y ≤ 1. S. Pollock Texas A&M Mathematics 12/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Numerics: P2- mesh Z κ(u )∇u · ∇v = 1 Z κ(u ) = 1 + −2 10 + (u − 0.8)2 g (x , y ) = 200 exp −200 (x − 0.7)2 + (y − 0.7)2 Finite element mesh fv , Finite element mesh Finite element mesh Figure: Mesh after 6 adaptive refinements. Left: PR+DR; center: DWR; right: PR + DWR S. Pollock Texas A&M Mathematics 13/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Numerics: P2- global error The maximum value of uh (x , y ) should approach 1. ITER PR + DR DWR PR PR +DWR 2 4 6 8 10 12 1.01016 1.00206 1.00101 0.999872 1.00016 0.999966 1.00333 1.00846 1.00138 0.999949 0.999065 0.999117 1.01008 1.00184 1.00111 0.999765 1.00001 0.999918 1.01032 1.00344 1.00158 0.999784 0.999941 0.999967 Properties that may indicate necessary refinement for the primal problem: quasilinear problems or stronger nonlinearities near-singularities in problem data: nontrivial quadrature error smooth solution with global support on domain S. Pollock Texas A&M Mathematics 14/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014 Main references Becker, R. and Rannacher, R., An optimal control approach to a posteriori error estimation in finite element methods, Acta Numerica,10, 1-102,2001 Holst, M., MCLite: An Adaptive Multilevel Finite Element MATLAB Package for Scalar Nonlinear Elliptic Equations in the Plane, UCSD Technical report and guide to the MCLite software package. Holst, M., Tsogtgerel, G., and Zhu, Y. Local and Global Convergence of Adaptive Methods for Nonlinear Partial Differential Equations, 2009. Holst, M., and Pollock, S., and Zhu, Y. Convergence of goal oriented methods for semilinear problems, 2012. S. Pollock Texas A&M Mathematics 15/15 GOAFEM for quasilinear problems Finite Element Circus October 24-25, 2014