Language models suggest ingredients and firing temperatures for inorganic materials recipes

The work

Technical · from the record

Seven off-the-shelf language models were prompted with 40 in-context examples to recommend precursors and predict calcination and sintering temperatures for inorganic solid-state synthesis, using held-out 1,000-entry subsets of a text-mined literature dataset. GPT-4.1 reached 53.8% Top-1 exact-match accuracy on precursor prediction, and reported mean absolute errors were 126 °C for calcination and 101 °C for sintering temperatures; averaging or rank-fusing three models improved results beyond Top-1 and cut inference cost by around 70% relative to GPT-4.1. The authors then used the models to produce 28,548 complete synthetic recipes and trained SyntMTE, a transformer derived from the DFT-pretrained MTEncoder, first on the synthetic data and then on literature data; a model trained only on synthetic data was 6% worse than the literature-trained model, while two-stage training improved mean absolute error by over 4% for SyntMTE. Applied to 40 reported routes for doped Li7La3Zr2O12 with all LLZO compounds withheld from training, the model reproduced the reported ordering of dopant-dependent sintering temperatures, including the drop for Bi substitution.

How AI was used

A text-mined inorganic synthesis dataset was filtered and split, with two 1,000-entry subsets held out and 40 in-context examples drawn from the validation fraction. Seven language models were queried through the OpenRouter API at temperature 0.1 with a parser-aware retry policy: one task asked each model to generate ranked precursor sets without being told how many precursors to use, with outputs canonicalised via pymatgen and scored by Top-k exact match; the other asked for calcination and sintering temperatures as structured JSON, scored by MAE, RMSE and R². Predictions from three-model ensembles were then combined without further learning, by min-, average- and max-rank fusion for precursor lists and by averaging for temperatures. Separately, lab-synthesized compounds from the Materials Project were reduced by maximum-entropy sampling over MTEncoder representations, non-solid-state entries were flagged out by GPT-4.1, and precursor sets and processing conditions were generated for the remainder, with incomplete routes and sub-threshold temperatures removed. SyntMTE encodes the target and all precursors with shared MTEncoder weights pretrained on the Alexandria DFT database, mean-pools and concatenates the embeddings, and regresses both temperatures through a two-layer MLP head; it and three baselines were trained under synthetic-only, literature-only and sequential synthetic-then-literature regimes on a chronological split. For the case study, an o3 model extracted temperatures from an LLZO literature corpus, LLZO and commonly doped variants were excluded from training and validation, and the fine-tuned model was applied to the extracted compositions.