[Proceedings of the AISB workshop on Evolutionary Computing, pp 34-48, University of Brighton, UK, April 1996 ]

Evolutionary design of synthetic routes in chemistry

Hugh M. Cartwright¹ and Julie A. Hopkins²

¹Physical and Theoretical Chemistry Laboratory, Oxford University,

South Parks Road, Oxford OX1 3QZ, UK

²Defence Research Agency, St. Andrews Road, Malvern WR14 3PS, UK

hugh@muriel.pcl.ox.ac.uk julie@muriel.pcl.ox.ac.uk

1 Introduction

The development of efficient methods to synthesise chemicals is a crucial task for the chemical industry. New synthetic routes were once developed largely by analogy with known reactions, but more systematic approaches are now used, in which computers play a central role.

Most computer-driven synthesis programs incorporate a database of chemical reactions and expert system rules. These are used to propose a reaction sequence which transforms readily-available starting materials into the desired product. Very many possible routes for such transformations may exist, depending upon the number of intermediates and the complexity of the product, and this presents synthesis design programs with various difficulties. For example, expert system rules must be numerous and precise. Furthermore, synthesis design operates in a large search space which, because of its discrete nature, is not smooth or differentiable. Traditional algorithms are adequate for small-scale problems, but are of limited use when the problem becomes large.

Since determination of the route of highest yield is an optimisation problem, it is reasonable to suppose that evolutionary algorithms might be of value in searching these large spaces. Genetic algorithms¹, spin glass models², connectionist architectures³, reaction diffusion systems⁴ and simulated annealing⁵ are already widely used in science, and synthesis design is a promising candidate for attack by an evolutionary algorithm.