Development of multimodal AI for photobiorefineries via knowledge syntheses, transfer learning, and techno-economic analysis

Synechococcus elongatus UTEX 2973 has emerged as a promising phototrophic chassis for next-generation biorefineries. However, the development of new cyanobacterial platforms is hindered by fragmented datasets, limited strain knowledge, and inadequate modeling tools. To address this challenge, we present MAP (Multimodal AI for Photobiorefinery), an integrated framework that combines large language models for knowledge synthesis, transfer learning for biological prediction, and techno-economic simulations for process evaluation. MAP employs GPT-4 and Graph-RAG (retrieval-augmented generation) to synthesize fragmented literature into coherent datasets and enable trustworthy knowledge retrieval. It further employs cross-species transfer learning from Synechocystis sp. PCC 6803 and S. elongatus PCC 7942 to improve the prediction power for growth and production traits of the emerging UTEX 2973 strain (R² > 0.93). Beyond data mining and performance prediction, MAP integrates model outputs with techno-economic analysis (TEA), enabling scenario evaluation across reactor types, cultivation conditions, and product pathways. Case studies in lycopene and sucrose production highlight how MAP pinpoints cost drivers, informs strain engineering, and guides process optimization. Ultimately, this multimodal workflow bridges biological performance with process economics and accelerates the design-build-test-learn (DBTL) cycle for data-driven photobiorefineries.