Wesley Y. Chen and colleagues use language models to turn sol-gel reports into roughly 148,000 synthesis records, including about 28,000 with impurity phases. Ingredients, conditions and outcomes remain linked, making competing products searchable.[1][2]
Sol-gel synthesis begins with liquid ingredients that form a gel. For bismuth ferrite films, chemical ratios and heating help determine which compounds emerge. Kevin Cruse and colleagues used manually collected recipes to guide new experiments in a 2024 study.[3]
One replication series shows why the detail matters. A published procedure omitted the bismuth-to-iron ratio. With other selected conditions held fixed, the team obtained an amorphous film at 0.9:1, an impurity at 1:1 and phase-pure bismuth ferrite at 1.05:1. One intended material yielded three outcomes.[3]
We reran the authors’ public validation code and labels: 51 of 104 papers met the strict rule of every evaluated field correct and one record per experimental variant. Individual facts can be accurate while recipes merge.[2]
Useful searches reveal experimental contrasts: the same target, a changed condition, a different outcome. A recipe’s structure becomes part of the scientific evidence.[3]
Sources & further reading
- Text-mined dataset of sol-gel synthesis protocols with phase purity outcomes using large language models Scientific Data · 15 Oct 2026 · Accessed 29 Oct 2026
- Sol-gel data hackingmaterials · Accessed 29 Oct 2026
- Text Mining the Literature to Inform Experiments and Rationalize Impurity Phase Formation for BiFeO3 Chemistry of Materials · 23 Jan 2024 · Accessed 29 Oct 2026