Applying Differential Transforms and ADER to Multi-Dimensional Atmospheric Transport and Non-Linear Dynamics

Differential Transforms (DTs), a core component of so-called "automatic" or "algorithmic" differentiation, offer significant flexibility and efficiency to any numerical method. The i-th and j-th DT, U(i,j), of a function, u(x,y), is simply U(i,j)=1/(i!j!)*∂^(i+j)u/∂xⁱ∂y^j. Being a term in the Taylor series of u(x,y) makes the reverse transform trivial. This relation also computes initial DTs from known spatial derivatives. What is novel about DTs is how they simplify a complex PDE system, transforming most arithmetic, trigonometric, and other operators into simple recurrence relations in derivative space. This allows one to simply and quickly compute analytical derivatives of highly complex and non-linear functions. Consider a pseudo-conservation law system, u(x)_t+f(u,x)_x=s(u,x), for instance. The fluxes and source terms could be (and often are) highly complex, non-linear functions of the state vector and independent variables. Regardless of the spatial discretization (variational / finite-element, weak / finite-volume, or strong / finite-difference), one nearly always must resort to tensored quadrature to evaluate face fluxes and body source terms, and this treatment is expensive. However, if one uses DTs to analytically compute spatial derivatives of the flux and source terms, given spatial derivatives of u, then the fluxes and source terms are directly expanded as polynomials, allowing for significantly cheaper, quadrature-free integration, sampling, and differentiation with a single dot product. Besides being simpler, this also allows flexibility for Galerkin methods in particular to analytically and cheaply compute body integrals, which are often approximated inexactly with quadrature. Computing Nth-order DTs in D dimensions is of O(D^2*N) complexity, and whether for transport or non-linear compressible Euler equations, they are cheaper to compute and integrate analytically than quadrature. Further, because time-dependent PDE systems relate spatial derivatives to time and space-time derivatives, given a set of known spatial derivatives, one can use DTs analytically compute time and space-time derivatives cheaply. Then, integrating directly in time over these space-time expansions of the PDE terms creates a time discretization method of the same philosophy as ADER methods but with significantly less expense than the explicit Cauchy-Kowalewski procedure, which can be exponential in complexity with respect to temporal order of accuracy. Because DTs compute mixed-dimension spatial derivatives and mixed space-time derivatives, over a single time step, all terms of the PDE are fully coupled to arbitrarily high-order accuracy in all spatial dimensions and time over a time step. DT-based ADER time discretizations lead to fully coupled, non-linear, accurate time stepping without resorting to multiple stages such as Runge-Kutta methods. This is advantageous in massively parallel computing environments because communication frequency and volume are reduced, leading to greater parallel efficiency. Results in two and three dimensions for linear transport and non-linear compressible Euler equations will be presented as well as accuracy and timing measurements to assess the efficiency for atmospheric models of applying DTs at various levels within Finite-Volume and Finite-Element discretizations of the underlying PDEs.