Large language model-assisted design of electrospun bioplastics

Abstract

Addressing the dual challenges of plastic waste and materials efficiency is critical for advancing sustainability. Here, a biodegradable electrospun polymer formulation was identified through a domain-adapted large language model (LLM) workflow and examined experimentally as an initial candidate in the search for sustainable alternatives to conventional plastics. The full text of 4022 research articles, comprising ∼8.6 million tokens, was processed locally with LangChain and ChromaDB to create a vector database for retrieval-augmented generation (RAG), while the LLM was separately fine-tuned using 200 domain-specific question-answer pairs. Guided by prompt-engineering strategies, the model suggested multiple biodegradable polymer formulations, from which a polylactic acid-polycaprolactone blend in tetrahydrofuran was selected for experimental validation. Electrospun fibers produced from this blend were characterized for chemical structure, morphology, and mechanical performance, revealing smooth and branched fibers with stable chemical bonds, while tensile testing showed that the mechanical strength of the as-spun fiber mat without any post-processing was lower than that of commercial polyethylene terephthalate (PET). These results present an initial proof-of-concept of a literature-guided workflow for identifying experimentally feasible biodegradable polymer formulations and supporting first-stage laboratory validation.