COntext-informed AUtomaTic cHemical prOcess geneRation

Improving the efficiency and resilience of the chemical and pharmaceutical industry is pivotal for the EU; in this frame, process flow diagram (PFD) design and retrofitting play a crucial role. Currently, it is executed by engineers leveraging their knowledge and engineering context, using a time-consuming iterative approach. On the other hand, AI-driven PFD design has proven outstanding capabilities in generating new process configurations. Despite being a promising technology, current approaches rely on training a new model when designing a process without leveraging process engineering information. Inserting such information would give the agent the optimisation context, improving the efficiency of the generated PFD and removing the need for retraining on each new design. However, such an integration requires investigating the best approach to inform the agent in the engineering context. This action wants to fill this gap by using a multidisciplinary approach combining reinforcement learning (RL) with chemical process optimisation and chemical properties predictions. The main project outcome is an AI agent able to design PFDs and retrofit existing ones without retraining while informed of the design aim and engineering context. The project will investigate the most suitable techniques for integrating chemical components and reaction properties into RL training for PFD design and retrofitting. The results of CoAUTHOR will advance our understanding of context-aware RL for chemical applications and generate an extensive toolkit for AI-driven PFD generation, paving the way toward a more sustainable and resilient chemical industry. In this action, I will combine my background in AI application in the chemical industrial domain with the expertise of the host research group in AI-driven PFD generation at TU Delft while acquiring skills in representing physicochemical properties via AI techniques during my secondment at ETH.