A strategy for the automatic synthesis of plant operating procedures, through the integration of artificial intelligence and mathematical modeling tools is discussed. This "predicative/numerical" modeling
IntroductionThe chemical process industry is increasingly interested in relatively small-scale, high-value, short-life-cycle products. This shift in industry perspective, justifies a fresh look at the way processing equipment is designed, programmed, and operated in order to ensure its flexibility. Flexibility of operation allows for the manufacture of different products using the same basic equipment. It also prolongs the useful life of a production facility beyond the short life cycle of a particular product. Exploiting the inherent flexibility of such plants gives rise to difficult operation problems. Although programmable logic controllers provide effective automation and sequencing of tasks, they are "hardwired" in the sense that they must be reprogrammed for each new production cycle. This raises the cost of flexibility and limits their effective capacity to adjust to plant upsets. In addition, as the size of the plant increases, the programming of these controllers become more involved. The automatic generation of batch plant procedures has therefore emerged as an important issue.The analysis and synthesis of process operations require models of the considered problem. The models developed for the operation of flexible production facilities should be flexible themselves (that is, they should be easily modified in order to mirror changes in the plant or manufacturing conditions). Traditional modeling techniques are not explicit enough to meet Correspondence concerning this article should be addressed to R. Lavie. this requirement, while those based on artificial intelligence show promise (Stephanopoulos et al., 1990;Piela et al., 1991). The need for explicit methods becomes apparent when the size of mathematical models grows and makes their development and understanding difficult. This is the case when we must deal with models consisting of thousands of variables and constraints.Recently, there has been interest in the integration of artificial intelligence (AI) and operations research (OR) techniques for the solution of MI(N)LP [mixed integer (non)linear programming] models. OR methods are useful for the efficient solution and analysis of numerical optimization problems. However, A1 methods may provide two main avenues of potential benefit: (1) they provide modeling methodologies to simplify the generation, correction, and reuse of the process OR models; (2) they may reduce the computational load by systematically exploiting the problems logical structure and the accumulated experience. Previous works were based on propositional logic and concentrated on aspects related with the second avenue. Two examples are the generation of propositional logic models from the design superstructures considered, or the use of the logic relationships to aid the branch and bound search in the solution of MI(N)LP proble...