Dynamical Analysis of the Snow Equation Simon Fraser University University of British Columbia Department of Mathematics Department of Physics Pouya Bastani Kevin Mitchell Ralf Wittenberg Tom Tiedje . – p.1/22 Outline Introduction The Model Numerical Schemes Numerical Results Future Work . – p.2/22 . – p.3/22 Introduction Suncups form on the surface of snow during the summer as a result of melting and evaporation due to solar radiation Two-dimensional patterns with a periodicity of 20-80 cm (characteristic length) and typically 2-50 cm deep Formed because hollows trap incident sunlight more efficiently and thus melt faster than peaks, leading to instability . – p.4/22 The so-called “Snow Equation” is a mathematical model of snow surface growth based on solar radiation, recently proposed by T. Tiedje et al. T. Tiedje, K.A. Mitchell, B. Lau, E. Nodwell, Radiation Transport Model for Ablation Hollows on Snowfields , J. Geophys. Res. 111 (2006) Derived from the effect of the snow surface shape on the absorption and diffusion of light in snow Explains the shape, size, and dynamical behavior of suncups in terms of the interaction of solar radiation with snow . – p.5/22 The Model . – p.6/22 Assumptions h(x, y, t): surface height of snow relative to a horizontal reference plane ∂h ∂t ∝ rate of absorption of solar radiation (light-snow interaction) Assumptions: weak surface topography: |∇h| ≪ 1 vertical incident light Limitations: not accurate when suncups become deep neglects the fact that on sloping surfaces suncups tilt toward the sun . – p.7/22 Light-Snow Interaction Surface Topography: Slope dependence: light at off-normal incidence to the snow surface is more likely to be scattered out of snow Curvature dependence: for concave surfaces light must travel further through the snow to escape, and so more of it is absorebed. Multiple reflections: radiation intensity increases proportionally to the solid angle of snow subtended by the surrounding snow and its albedo(The fraction of light scattered back into the atmosphere) Flow of radiation in the snow . – p.8/22 The Snow Equation Non-dimensionalized snow equation: ∂t h = −∇2 h − ∇4 h − α|∇h|2 + β∇2 |∇h|2 α2 + β 2 = 1 ∇ is the Gradient, ∇2 is the Laplacian, ∇4 is the Biharmonic operator Snow equation in one dimensional periodic domain ht = −hxx − hxxxx − α(hx )2 + β(hx )2xx x ∈ [0, L] Similar equations were developed in other surface growth models by T. Tiedje et al. at MBE Lab, UBC . – p.9/22 Kuramoto-Sivashinsky Equation β = 0 gives the (integrated) Kuramoto-Sivashinsky (KS) Equation ht = −hxx − hxxxx − (hx )2 Contexts: flame fronts, plasma ion waves, chemical phase turbulence, ... An example of a deterministic dynamical system that exhibits complex spatiotemporal phenomena and chaotic behaviour . – p.10/22 Fourier Space Evolution On a periodic domain [0, L], Fourier representation of u(x, t) is: h(x, t) = ∞ X k=−∞ b hk (t)eikx 1 b hk (t) = L Z L h(x, t)e−ikx dx k= 0 2nπ L P ′ 2 4 b b hk′ b hk−k′ KS in Fourier space: ∂t hk = (k − k )hk − k′ k (k − k ′ )b Infinite system of ordinary differential equations hxx term destabilizes large scales, hxxxx damps small-scale modes, hhx provides nonlinear coupling between scales Dispersion relation: γ = k 2 − k 4 , reality condition: b hk = b h∗−k . – p.11/22 Nonlinearity in real space ⇒ coupling of modes in Fourier space Solutions with fundamental period less than L are unstable Power Spectrum: small modes (large scales): flat power spectrum active scales: modes near the peak of the spectrum large modes (small scales): decay rapidly KS Power Spectrum 1 10 0 10 −10 10 0 |v|2 10 −20 10 −1 10 −30 −1 0 10 1 10 k 10 |v|2 10 −2 10 5 −3 h 10 0 −4 10 −5 −50 −5 −40 −30 −20 −10 0 x 10 20 30 40 10 −2 10 −1 0 10 10 1 10 k . – p.12/22 Numerical Schemes . – p.13/22 Finite Differences (FD) Intuitive, and simple. Accuracy and numerical stability issues θ-method for Heat Equation: ut = uxx 2 n+1 n + (1 − θ)δx2 unj θδ − u un+1 x uj j j = ∆t ∆x2 δx2 unj = unj+1 − 2unj + unj−1 θ = 0: Euler Forward Difference Method - Explicit stable if ∆t ≤ 0.5∆x2 , error = O(∆t + ∆x2 ) θ = 0.5: Crank-Nicolson Method - Implicit unconditionally stable, error = O(∆t2 + ∆x2 ) θ = 1: Backward Difference Method - Implicit unconditionally stable, error = O(∆t + ∆x2 ) . – p.14/22 Spectral Methods Periodic discrete real domain ⇔ discrete periodic Fourier domain Discrete Fourier Transform (DFT) and Inverse DFT: Fk {u} = u bk = ∆x N X −ikxj uj e u} Fj−1 {b j=1 xj = j∆x 1 = uj = 2π N/2 X k=−N/2+1 u bk eikxj x ∈ [0, 2π] Real space differentiation ⇔ Fourier space multiplication: ub′ (k) = ikb uk Fast Fourier Transform (FFT): O(N log N ) instead of O(N 2 ), N = 2n . – p.15/22 Integrating Factor (IF) The snow equation is a “stiff" PDE requiring extremely small time stepping to guarantee stability with FD Methods Stiff because the magnitude of the fourth-order linear term is large in the Fourier space for large modes: u bt = −k 4 u b + ... Transform the PDE to solve the linear part exactly bt = −(α + βk 2 )e−(k2 −k4 )t F {F −1 {ike(k2 −k4 )t U b }2 } U Treat the nonlinear terms in real space b = e−k2 t u U b Allows 5 or 10 times larger time steps than FDM’s . – p.16/22 Exponential Time Differencing (ETD) S. M. Cox, P. C. Matthews, Exponential Time Differencing for Stiff Systems, J. Comp. Phys. 176 (2002), pp. 430-455 Developed the ETDRK4 (Exponential Time Differencing fourth-order Runge-Kutta) method for solving stiff nonlinear PDEs A.K. Kassam, L.N. Trefethen, Fourth-Order Time-Stepping For Stiff PDEs, SIAM J. Sci. Comp. 26 (2005), pp. 1214-1233 Solved the problem of numerical instabilities and generalized to nondiagonal operators Modifications to the implementation by Kassam and Trefethen . – p.17/22 Adaptive Time Stepping Achieve a predetermined accuracy with efficient time stepping Calculate two independent solutions of the next time step by independently taking one whole step and also two half steps Estimate the error using these two solutions Adapt the stepsize: if error < threshold increase time step accordingly and continue to the next time step otherwise decrease time step and repeat until desired accuracy is achieved . – p.18/22 Results and Observations . – p.19/22 Numerical Results ht = −hxx − hxxxx − cos θ (hx )2 + sin θ (hx )2xx π θ ∈ [0, ] 2 x ∈ [−16π, 16π] θ = π/4 θ = π/2 . – p.20/22 Observations Stationary or moving patterns, depending on the parameter θ Characteristic length and amplitude depend on θ 14.5 14 13.5 13 char. len. 12.5 12 11.5 11 10.5 10 9.5 0 0.05 0.1 0.15 0.2 θ 0.25 0.3 0.35 0.4 0.45 Need larger domain size, or equivalently, smaller Fourier mode spacing, to get more accurate results . – p.21/22 Future Work Validity of the model Generalization of the model to any angle of incident light Field observations and comparison with numerical solutions Further calculations of dynamical variables, such as characteristic, amplitude, Lyapunov exponent, ... . – p.22/22