Sprecher
Beschreibung
Engineering problems in sustainable process design are often introduced as broad stakeholder objectives, such as improving efficiency, reducing costs, and enhancing environmental performance. However, these objectives may not sufficiently describe system boundaries, causal relationships, technical constraints, and engineering contradictions. This can limit the effectiveness of subsequent idea generation and concept development.
This contribution presents a multi-role agentic AI framework for systematic analysis of engineering problems. The framework uses four specialised AI roles that sequentially perform problem framing, cause-and-effect analysis, contradiction identification, and problem prioritisation, thereby transforming broad stakeholder objectives into a structured engineering problem formulation. The framework was implemented using GPT-5.5 and benchmarked against a conventional single-role baseline using a biodiesel production case study focused on the transesterification stage. In this process, feedstock variability, catalyst efficiency, operational costs, and sustainability trade-offs create substantial analytical complexity. Both configurations were evaluated using a hybrid AI-human expert approach across four criteria: problem-framing completeness, causal reasoning quality, contradiction identification, and stakeholder-objective alignment.
Across five experimental runs per configuration, the multi-role framework significantly outperformed the single-role baseline across all criteria, achieving an overall score of 4.60 compared with 2.82 out of 5 (p = 0.031). The largest improvements were observed in contradiction identification and causal reasoning quality. Notably, the multi-role framework identified approximately three times as many engineering contradictions per run and more consistently captured both technical and environmental trade-offs.
These findings suggest that decomposing engineering problem analysis into role-differentiated analytical stages produces more systematic and comprehensive problem formulations than conventional single-role LLM interactions. The results demonstrate the potential of agentic AI to support early-stage sustainable process design.
