Integrated Retrosynthesis and Large Language Modeling for Chemistry-Informed Circular Chemical Reaction Networks

Discovery and development of manufacturing routes that explicitly account for the entire product life cycle are essential for the transformation to a sustainable and circular chemical industry. Retrosynthesis is a promising approach, but operates within a gate-to-gate paradigm, limiting its ability to explicitly integrate end-of-life waste streams into upstream production pathways. We evaluated four reaction discovery strategies that combine human intervention, pattern recognition, retrosynthesis, and large language models (LLM) within a common comparative framework. When evaluated with methanol as a benchmark system, the strategies reveal distinct trade-offs between pathway discovery and technological maturity. Combining LLMs with retrosynthesis achieves the broadest expansion of reaction space with 165 reactions involving 12 unique chemicals and the highest novelty relative to the conventional business-as-usual reference network. The resulting circular chemical reaction network (CCRN) for a more complex molecule, polyethylene, contains more than 1,000 reactions, and human-guided extraction increased the total number of identified reactions by 34–35%, depending on the keyword-search strategy, while recovering additional end-of-life reactions from experimental results. The proposed framework demonstrates that combining language-based knowledge extraction with structure-informed retrosynthetic reasoning enables scalable construction of CCRNs and supports the systematic exploration of candidate circular chemical pathways for subsequent economic and environmental evaluation.