From analytic to automated derivatives: a case study of the electrostatic potential.

Abstract

Given a large-scale engineering simulation, one is often interested in the derivatives of certain program outputs with respect to certain input parameters, e.g., for sensitivity analysis. From an abstract point of view, the computer program defines a mathematical function whose derivatives are sought. In this note, two cases are investigated where the underlying functions are given by the solution of different two-dimensional electrostatic potential problems. For a rectangular region, the function and its derivatives can be derived analytically. For a function based on an L-shaped region, the function is computed by a simulation code and its derivatives are obtained by automatic differentiation. In general, this powerful technique is applicable if a function is given in the form of a computer program in a high-level programming language such as Fortran, C, or C++. In contrast to numerical differentiation providing approximations based on divided differences, the derivatives computed by automatic differentiation are accurate. Furthermore, automatic differentiation is also shown to be computationally efficient.