A perfectly matched layer approach to a reactive scattering problem Introduction The nuclear TDSE Numerical challenges Numerical methods Anna Nissen1 , Hans O. Karlsson2 and Gunilla Kreiss1 Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching 1 2 Division of Scientific Computing, Department of Information Technology, Uppsala, Sweden. Quantum Chemistry, Department of Physical and Analytical Chemistry, Uppsala, Sweden. Numerical experiments Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (1 : 31) Motivation and background Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments (a) Reaction rates (b) Femtosecond laser , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (2 : 31) The nuclear Schrödinger equation We consider the nuclear time-dependent Schrödinger equation Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching i~ ∂ ψ(X, t) = H(X)ψ(X, t), ∂t where the influence from the electrons has been modelled into a potential energy surface V(X) through the Born-Oppenheimer approximation. The Hamiltonian is given by Numerical experiments H(X) = T + V = X i − ~2 ∇2N,i 2Mi + V (X). X, Mi – Nucleus coordinates andRmass on nucleus i. 2 |ψ|2 – Probability distribution, Ω |ψ| dΩ = 1. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (3 : 31) Numerical challenges I High-dimensionality (3N-6 DOF for N atoms). Example: molecule with 5 atoms ⇒ 9 DOF. Introduction The nuclear TDSE Numerical challenges Numerical methods I Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Highly oscillatory solutions. Couplings to strong electromagnetic fields. I Restrict computational domain to a finite domain (dissociative problems). Numerical experiments Two approaches: Absorbing boundary conditions and absorbing layers. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (4 : 31) Numerical methods Introduction The nuclear TDSE Numerical challenges Numerical methods I High order finite difference methods in space. I Arnoldi algorithm in time, corresponding to the Lanczos algorithm for general matrices. I Currently Matlab implementation ⇒ MPI and OpenMP parallelized code HAParaNDA [1]. Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments [1]. M. Gustafsson and S. Holmgren, Proceedings of ENUMATH 2009, Uppsala, Sweden. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (5 : 31) Absorbing layers Extension of the computational domain. Introduction Interior domain [0,x0 ] Absorbing layer [x0 ,x0 + d] The nuclear TDSE Numerical challenges Numerical methods 0 x0 x0 + d Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching I Modify the potential operator V ⇒ V − iW . Complex absorbing potential (CAP). I Modify the kinetic operator T through a coordinate transformation, x ⇒ F (x). Exterior complex scaling (ECS), Perfectly matched layer (PML). Numerical experiments , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (6 : 31) Absorbing layers Introduction V ⇒ V − iW I Complex absorbing potential: Ref. [2], [3]. I Complex scaling: x ⇒ F (x) = xe iθ Ref. [4]. The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching [2] D. Neuhauser, M. Baer, J. Chem. Phys. 119:77-89, 2003. Numerical experiments [3] U. V. Riss, H-D. Meyer, J. Phys. B 26:4503-4535, 1993. [4] N. Moiseyev, Physics Reports 302:211-293, 1998. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (7 : 31) Absorbing layers I Exterior complex scaling (ECS) Ref. [5]. x, x < x0 F (x) = , x0 + e iγ (x − x0 ), x ≥ x0 I Smooth exterior scaling (SES) / Perfectly matched layer (PML) Ref. [6], [7]. x ⇒ F (x) after x = x0 , where dF dx is cont. at x0 for a smooth transition. Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments [5] C. W. McCurdy, C. K. Stroud, M. K. Wisinski, Phys. Rev. A 443:5980-5990, 1991. [6] H. O. Karlsson, J. Chem. Phys. 109:9366-9371, 1998. [7] J. P. Berenger J. Comput. Phys. 114:185-200, 1994. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (8 : 31) Perfectly matched layer 2D Schrödinger equation with a potential ∂ψ ∂2ψ ∂2ψ =− 2 − + V (y ) in Ω = [0, ∞) × [−L, L], ∂t ∂x ∂y 2 ∂2 2 j = 1, ..., ∞. where − ∂y 2 + V (y ) ψj = κj ψj Assume compactly supported initial conditions ψ0 (x), homogeneous boundary conditions at y = ±L, a Dirichlet condition at x = 0 i Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching ψ(0, y , t) = 0, Numerical experiments and a decay condition lim ψ(x, y , t) = 0. x→∞ Want to reduce the computational domain to Ω = [0, x0 + d] × [−L, L]. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (9 : 31) Find bounded solutions Laplace-transform outside compact support of initial data is ψ̂j = − Introduction The nuclear TDSE Numerical challenges ∂ 2 ψ̂j + κ2j ψ̂j ∂x 2 in Ω = [x0 , ∞) × [−L, L] yields ODE in x with modal solutions Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments ψ̂j (x) = Aj e λ+ (x−x0 ) + Bj e λ− (x−x0 ) , q λ± = ± −is + κ2j . (1) Consider Re(s) ≥ 0 and choose q −is + κ2j ≥ 0 for Re(s) ≥ 0, Re ψ̂j (x) = Bj e − q −is+κ2j (x−x0 ) . , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (10 : 31) Modify solutions Substitute q −is + κ2j (x − x0 ) against Z x q iγ 2 −is + κj ((x − x0 ) + e σ(ω)dω) Introduction The nuclear TDSE Numerical challenges x0 Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments in the PML. Absorption function σ(ω) – smooth, non-negative function in ω. σ(x0 ) = 0 for perfect matching. The modified solutions look like ψ̂j PML (x) = Bj e − q R −is+κ2j ((x−x0 )+e iγ xx σ(ω)dω) 0 , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo . (11 : 31) PML equation Introduction The nuclear TDSE Numerical challenges Transforming modal PML solutions back into physical space yields PML equation 2 1 ∂ 1 ∂ψ ∂ ψ ∂ψ =− − 2 +V (y )ψ, i iγ iγ ∂t 1 + e σ(x) ∂x 1 + e σ(x) ∂x ∂y Numerical methods Absorbing layers in Ω = [0, x0 + d] × [−L, L]. Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments Can be viewed as a coordinate change from real coordinate x in the interior domain to complex coordinate F (x) in the layer, where Z x F (x) = x + e iγ σ(ω)dω. x0 , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (12 : 31) Well-posedness Introducing ϕ = i Introduction f ψ yields ∂ϕ 1 ∂2 1 =− ϕ + Vk (x)ϕ + V (y )ϕ, ∂t f (x) ∂x 2 f (x) The nuclear TDSE Numerical challenges Numerical methods √ where Vk (x) = 3f 2 −2f 00 f 4f 4 , f (x) = dF (x) dx (2) = 1 + e iγ σ(x). Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments Well-posedness for (2) for 0≤γ≤ π , 2 follows from the bound ||ϕ(·, t)||2H p ≤ Kp (T )||ϕ(·, 0)||2H p , where Kp depends on derivatives of f up to order p for p ≥ 2, and up to order 2 for p = 0. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (13 : 31) Continuity requirements Introduction The nuclear TDSE Numerical challenges Finite difference discretization in space of order m yields 1D error equation ( 1 ∂2 1 iet = − f (x) e + Vk e + ∆x m Te , ∂x 2 f (x) e(x, 0) = 0, Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching where ej (t) = ϕ(xj , t) − uj (t). Here m is assumed to be even and cm ∂ m+2 1 ϕ . Te = f ∂x m+2 f Numerical experiments Thus, the bound ||ϕ(·, t)||2H m+2 ≤ Km+2 (T )||ϕ(·, 0)||2H m+2 is needed for optimal convergence, implying that f should have m + 2 bounded derivatives. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (14 : 31) Illustration of PML errors Introduction Two main types of errors from the PML: The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments I Modeling error – due to solving equation on truncated domain [0,x0 + d]. I Numerical reflections – discretization destroys perfect matching at interface x0 . I MOVIE , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (15 : 31) Modeling error Introduction I Error due to truncation of the computational domain. I Can be estimated in terms of the modal PML solutions in Laplace-space. I Modeling error ∝ e I k can be estimated as the dominating wavenumber of a wave packet. The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments −k Rx 0 +d x0 σ(ω)dω . , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (16 : 31) Numerical reflections Introduction I Depends on the discretization and of the absorption function σ(x). I High order central finite difference methods, order m (m = 8 typically). I Truncation error Te ∝ The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching cm ∂ (m+2) f ∂x (m+2) 1 fϕ . Numerical experiments , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (17 : 31) Numerical reflections Introduction I Largest effect from the vicinity of the interface ⇒ Te ∝ f α f (m+2) ϕ. I For polynomial profile of order p, p = m + 2, σmax p σ(x) = σdmax p (x − x0 ) , Te ∝ d p . I Smoothing effect since equation is parabolic in the PML (gain at least one derivative). The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (18 : 31) Error matching Explicit error formulas for polynomial absorption profile p σ(x) = σdmax p (x − x0 ) . Goal: For error level ε Introduction The nuclear TDSE Numerical challenges modeling error ε1 ≈ e −2ksin(γ)σmax d p+1 ≤ ε, Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching m ≤ ε. numerical reflections ε2 ≈ C σdmax p ∆x Procedure: I Given ∆x and error level ε. I Determine C from numerical reflections numerically, |ε2 | = ε gives optimal M = σmax /d p . I |ε1 | = ε gives optimal width d. I Optimal σmax determined from σmax = Md p . Numerical experiments , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (19 : 31) H2 molecule on a solid surface Introduction The nuclear TDSE Numerical challenges Numerical methods Dissociative adsorbtion (H2 molecule dissociates along the surface) and associative desorption (H2 molecule leaves surface). r - Internuclear distance in H2 molecule, z - Distance between molecule and surface. Potential energy surface in [8]. Absorbing layers Perfectly matched layer Derive PML equation Well-posedness 10 8 r PML errors Modeling error Numerical reflections Error matching 6 Numerical experiments 4 2 2 4 6 z 8 10 [8] R. C. Mowrey, . Chem. Phys. 94:7098-7105, 1991. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (20 : 31) Comparison of absorbing layers Numerical reflections for different absorbing layers (dissociative adsorbtion). 0 Introduction 10 PML CAP TCAP ECS 8th order The nuclear TDSE Numerical challenges Numerical methods Absorbing layers PML errors Modeling error Numerical reflections Error matching −5 10 l2 error Perfectly matched layer Derive PML equation Well-posedness −10 10 Numerical experiments −15 10 2 10 Number of points along r−axis , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (21 : 31) Comparison of absorbing layers Simulations with PML parameters optimized for certain tolerance TOL (dissociative adsorbtion). −1 Introduction 10 The nuclear TDSE Numerical challenges −2 Numerical methods 10 Absorbing layers PML errors Modeling error Numerical reflections Error matching −3 l2 error Perfectly matched layer Derive PML equation Well-posedness 10 −4 10 PML CAP TCAP TOL Numerical experiments −5 10 −6 10 −5 10 −4 10 −3 Tolerance 10 −2 10 , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (22 : 31) Outgoing flux Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments The conservation law for probability in quantum mechanics is given by Z Z ∂ 2 |Ψ| dΩ + j · dS = 0, ∂t Ω S (2) where S is the boundary of Ω, and the definition of the flux in direction x, jx , is ~ ∗ ∂Ψ = Ψ . (3) jx = mx ∂x Integrating equation (2) with respect to time yields Z |Ψ|2 dΩ + Ω since R Ω |Ψ0 | 2 dΩ Z T Z j · dS = 1, 0 (4) S = 1. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (23 : 31) Outgoing flux Introduction The nuclear TDSE Numerical challenges After a sufficiently long time T , Z T Z j · dS = 1, Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments 0 (5) S when all probability density has vanished from the computational domain. Equation (5) should be monotonously increasing, and this can be used as a measure of the accuracy of the boundary conditions. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (24 : 31) Discrete flux An analogue expression to the flux (3) which is exact in space for the discrete approximation can be derived from the semi-discrete Schrödinger equation in 1D Introduction The nuclear TDSE Numerical challenges i~ Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching ∂uj (t) ~2 =− Dm,j uj (t). ∂t 2M Dm,j – mth order central finite difference operator corresponding to a second derivative. uj (t) – semi-discrete solution in grid point xj . Numerical experiments Dm,j uj = 1 α0 uj + ∆x 2 m/2 X αk (uj+k + uj−k ) , k=1 where the α’s are the weights of the finite difference stencil. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (25 : 31) Discrete flux Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments A general expression for the discrete flux using a central finite difference discretization of order m, jDm (t), is derived, so that Z TZ Z T j · dS ≈ jDm (t)dt. 0 S 0 This expression is exact in space for the finite difference approximation and only the temporal integral needs to be approximated. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (26 : 31) Integrated flux 1 The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments Integrated flux Introduction 0.995 0.99 PML CAP TCAP 0.985 0.98 700 800 900 1000 Time 1100 1200 Figure: Integrated outgoing flux in the z-direction. The flux for the PML and the TCAP converge to one, while numerical reflections causes an unphysical decrease for the CAP flux. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (27 : 31) Integrated flux 1 Introduction Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments 1 Integrated flux The nuclear TDSE Numerical challenges 1 1 PML CAP TCAP 1 850 900 950 1000 1050 Time 1100 1150 1200 Figure: Integrated outgoing flux in the z-direction, closer view. Here it is visible that the flux decreases also for the TCAP. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (28 : 31) Integrated flux Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments Boundary condition reference PML optimized PML non-optimized PML CAP TCAP fluxz 0.011364 0.011389 0.011392 0.084520 0.018261 fluxr 0.98863 0.98857 0.98860 0.91469 0.98134 Table: Integrated flux in r - and z-direction, corresponding to fractions of transmitted and reflected parts of the wave packet, respectively. PML, CAP and TCAP boundary conditions with the same number of points are used. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (29 : 31) Integrated flux 1 Introduction Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments 0.95 Integrated flux The nuclear TDSE Numerical challenges 0.9 CAP PML reference TCAP 0.85 0.8 0 0.5 1 1.5 Time 2 2.5 3 4 x 10 Figure: Integrated outgoing flux in the r -direction. The accuracy of the boundary conditions affect the results to a large extent. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (30 : 31) Concluding remarks I Accuracy of the absorbing boundary conditions may affect the measured quantity considerably. I Both continuous and discrete reflection properties need to be taken into account in the boundary treatment. I Systematic procedure of choosing parameters saves time and leads to efficient absorbing boundary treatment. [9] Introduction The nuclear TDSE Numerical challenges Numerical methods Absorbing layers Perfectly matched layer Derive PML equation Well-posedness PML errors Modeling error Numerical reflections Error matching Numerical experiments [9]. A. Nissen, G. Kreiss, An optimized perfectly matched layer for the Schrödinger equation, To appear in Communications in Computational Physics. , Time dependent quantum mechanics – analysis and numerics, 28-30 April 2010, Oslo (31 : 31)