Share Email Print
cover

Proceedings Paper

Accurate neural network-based modeling for RF MEMS component synthesizing
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

Contrary to traditional analysis flows as expensive FEM simulation tools or inaccurate electrical models extractors, we developed MemsCompiler that implements a new real synthesis approach for RF MEMS. The new flow starts from system designer requirements and generates, in a one-click operation, a ready-to-fabricate layout (GDSII) and a passive fitted equivalent Spice circuit. Concerning the circuit, physical considerations give us an equivalent schematic in which circuit parameters values must be adjusted to fit the required performances. As to the GDSII, which constitutes the main contribution of this work, Design Of Experiment technique, used in the first version of the synthesizer, gave about 11% of dispersion and found to be unsatisfactory in some cases. A more accurate modeling was indispensable. Thus, we developed a neural networks-based modeling for circular inductors, which are considered by designers among the most stubborn components. This new modeling has shown to be very accurate: MemsCompiler produced about 3% of dispersion compared to the equivalent circuit and about 6% of dispersion for generated geometries. This modeling is flexible and could be rapidly generalized to other components.

Paper Details

Date Published: 24 January 2004
PDF: 8 pages
Proc. SPIE 5344, MEMS/MOEMS Components and Their Applications, (24 January 2004); doi: 10.1117/12.523954
Show Author Affiliations
Firas Mohamed, MEMSCAP SA (France)
Bachar Affour, MEMSCAP SA (France)


Published in SPIE Proceedings Vol. 5344:
MEMS/MOEMS Components and Their Applications
Siegfried W. Janson; Albert K. Henning, Editor(s)

© SPIE. Terms of Use
Back to Top