(456c) Semantically Enriched High-Throughput Optimization

The paper presents a systematic approach to represent and extract process synthesis knowledge in high-throughput optimization. Means of analysis include dynamic ontologies populated by computer experiments and continuously upgraded in the course of optimization. Process synthesis makes extensive use of superstructure optimization to address the systematic development of process design options through generic representations integrating all available designs. Although results can be produced in a single stage, practice dictates the employment of multiple stages - first with the deployment of simple conceptual models and then with the use of more detailed formulations ? as intermediate solutions can be reviewed, analyzed and understood. Even in cases where the development of solutions is rigorous and proven, synthesis solutions can prove rather inconclusive and difficult to translate. Indeed, process synthesis typically yields a multitude of solutions with (very) similar design performance but different structural and operational characteristics. In the same vane, minor changes in the design parameters often yield major changes in the selected designs (types of units, connections employed) which are difficult to interpret, comprehend or summarize. The systematic interpretation of the solution can yield not only to a better understanding of the solution space but also to a systematic reduction of the superstructure employed (both in size and complexity) and, thus, to much simpler experiments. The work combines optimization, semantic models (in the form of ontologies) and analytical tools. The research is presented against a background of stochastic optimisation techniques that enable the development of solution clusters for chemical reactors synthesis. The work explains the systematic extraction of information with a purpose to simplify and interpret design results. The simplification is achieved with a gradual evolution of the superstructure and corresponding adjustments of the optimization search. The interpretation is accomplished with the use of analytical tools to translate data into descriptive terms understood by users. Both stages make use of ontology models bridged with analytical tools to extract information patterns from solution data. Data mining and clustering techniques are employed in order to identify design trends (units, connectivities) and to build the components required for the knowledge models. Ontologies feature concepts and relationships processed by external analysis. They can be interfaced with synthesis workers and can be used to enable solution summaries or the visualization of the results. In the implementation, the work combines the optimization models with an ontology environment and a reasoner. Experiments are registered on the ontology that in turn guides the search of the analysis. The ontologies feature: (i) fixed concepts and relationships directly related to the structural and operational components of the superstructure and (ii) floating concepts upgraded with new associations and relationships in the course of analysis. Dynamic ontologies were developed using Protégé-OWL (supports OWL). They were supported by the RACER reasoner that was functioning as a logical classifier and an ontology verification tool. Visualization was partly enabled through the GrOWL editor. Results are presented in several examples of reactor network synthesis but are not essentially restricted to a particular type of application.