Poster (A0): Applying Automatic Differentiation to Computer Aided Process Engineering Software
by
Karim Alloula
CNRS LGC INPT ENSIGC
Coauthors: Jean-Pierre Belaud

Any CAPE software package is a collection of thermodynamic and unit operation models used to simulate chemical processes. It intensively uses derivatives in order to solve such models and get their sensitivities. In this area, up to now, process simulator designers face an alternative for obtaining function derivatives: writing derivatives by hand and/or getting their numerical approximation through a finite difference scheme.

The objective of this work is to evaluate several differentiation modes applied to thermodynamic and unit operation models coming from elf and BP in-house codes. The differentiation modes are hand-coded analytical derivatives, finite differences, Odyssée direct, reverse and Jacobian modes. This trial assesses quantitative and qualitative performance criteria. A strategy well-adapted to the application of AD is defined: 1- separate model function and model function derivative codes from the Newton-Raphson solving process; 2- replace model function derivative codes with AD generated ones; 3- solve Newton steps with a direct or iterative method from the total derivatives or from directional derivatives.

The derivatives generated by Odyssée prove to be particularly accurate. Odyssée's Jacobian mode coupled with a direct method shortens the thermodynamic model calculation time, while its direct mode coupled with an iterative method improves the unit operation model calculation time. Also, from a qualitative point of view, use of AD reduces the development and maintenance costs of CAPE codes.

These results are part of the work achieved for a European community ESPRIT project AD-CAPE. Thanks to all the partners especially Katerina Papalexandri (BP Amoco), Nicole Rostaing-Schmidt (Simulog), Olivier Baudouin (ProSim SA) and Michel Pons (elf Atochem).

Laboratoire de Génie Chimique de Toulouse - Process Analysis and Simulation

Date received: February 4, 2000