The chemical industry is positioning itself for a new era of digital research and development. The complexity of product development, along with the influx of information, entail a need for scalable and intelligent decision-making at different levels of the innovation value chain. This thesis tackles the problem of autonomous process design, by combining data-driven, knowledge-driven, and mathematical programming tools within a unified framework.

To realise the intelligent system, the process design problem is first reformu- lated as a reinforcement learning problem, wherein an agent can interact with an environment to build flowsheets, receive reward as feedback, and iteratively improve its designs. Further, a hierarchical agent is developed to first identify desired design tasks, then arrange units to achieve them, thus emulating an expert’s planning strategy. The agent is then combined with an ontological tool, i.e. a computer-readable knowledge base of processes, tasks, phenomena, and operational bottlenecks, along with their attributes and relations. This knowledge is processed using predicate logic to dynamically define states, actions, and objectives on the different design levels. The ontology also enables the combination of phenomena to attain intensified multi-functional unit operations.

This work represents a step-change in autonomous design by providing a novel end-to-end design procedure. Further development would enable continual, multi-source knowledge acquisition. This will ultimately lay the foundations for next-generation intelligent systems for scalable process design and intensification.