Behavioural synthesis of analogue integrated circuits

Abstract

Automatic synthesis of analogue circuits remains a very manually intensive task despite huge strides in the field of Electronic Design Automation (EDA) in recent decades. Genetic Algorithms (GAs) are biologically inspired search algorithms which have previously shown some promise in this field. Their ability to form the basis of a practically useful synthesis system is investigated. A GA-based experimental synthesis system is implemented, which employs a Genetic Programming (GP) style encoding scheme based on tree structures, and a novel fitness function based on pole-zero analysis. The system is capable of synthesising circuit topologies entirely from scratch, but can also utilise user-provided circuit knowledge of arbitrary detail and complexity. The system uses a SPICE-based circuit simulator as a circuit evaluator. Experimental results reveal a number of issues that adversely impact the ability of GAs to reliably synthesise practically useful analogue circuits. These include considerable resource requirements and a tendency for synthesised circuits to contain an unnecessarily large number of components. Most serious is the sensitivity of analogue circuits to changes in topology and/or sizing. GAs are shown to be currently ill-suited to the problem domain of analogue circuit synthesis. The problem of SPICE non-convergence on the GA is also considered.