Solving the seveso II prediction problem of substance generation during loss of control by Al techniques

The European Directive 96/82/EC forces industries to know exactly what substances may be generated during loss of control of an industrial chemical process. The study must be as accurate as possible, as the design of the safety measures that will be ready in case a real loss of control occurs depends on these results. Chemical reactions under out of control conditions, follow defined but difficult to know rules, specially at high temperatures showing high molecular complexity. It seems that the application of artificial intelligence techniques will help to predict which substances are produced given an initial chemical scenario. In this paper the k-nearest algorithm and its application to this problem will be studied. A database has been developed to manage reactive information independently from the algorithm or technique of AI to be applied. The most important information of the database are relations between substances present on the initial reactive scenario and the ones detected after reaction under out of control conditions occurred. Chemical substances contained in the database have been analysed by decomposing them into Benson's Groups, which is thought to be the most adequate level of detail to preserve chemical information while establishing relations between non-identical substances. The Shepard's modification of k-nearest neighbor algorithm has been used for the Boolean prediction of formation of a certain substance given a reactive scenario. Results show an accuracy above 95%.