The development of new materials increasingly depends on predictive models and multiscale simulations, yet essential data are often locked in unstructured scientific texts, limiting reuse. Conventional LLM‑based extraction typically yields flat key‑value pairs that do not respect ontology constraints. To address this, we introduce eolas, a modular pipeline that leverages large language models to automatically convert scientific documents into knowledge graphs aligned with a user‑specified ontology. eolas parses a full article in a few minutes and presents the resulting graph in a faceted table, facilitating quick human validation. We demonstrate the system on irradiated materials designed for the extreme temperature and radiation environments of fusion reactors, extracting entities such as composition, microstructure, experimental conditions, and performance metrics. Compared with a human expert who spends 30–90 minutes per paper, eolas achieves comparable quality in a fraction of the time. To benchmark LLM capabilities in ontology‑aligned KG construction, we release the first public dataset containing 168 experiments, each with source text, target ontology, and reference graph. Systematic experiments across various LLMs, prompting techniques, and post‑processing steps yield three practical guidelines: 1) Explicitly embed the ontology definition in prompts; 2) Apply a chunk‑extract‑merge workflow to reduce context overload; 3) Use structured validation scripts to filter inconsistent triples. Our results illustrate both the promise and current limits of LLMs for automated knowledge graph generation in materials science.
Review