Language Models Enable Data-Augmented Synthesis Planning for Inorganic Materials

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Inorganic synthesis planning traditionally uses heuristics or limited machine learning (ML) models, hindering broad applicability.
• Existing methods struggle with the vastness and complexity of chemical synthesis data.

Purpose of the Study:

• To evaluate the capability of off-the-shelf language models (LMs) for recalling and predicting inorganic synthesis conditions.
• To investigate the potential of LM-generated data for training more effective synthesis planning models.
• To develop a hybrid workflow for scalable and data-efficient inorganic synthesis planning.

Main Methods:

• Utilized pre-trained language models (GPT-4.1, Gemini 2.0 Flash, Llama 4 Maverick) for precursor and condition prediction.
• Ensembled multiple LMs to improve accuracy and reduce computational cost.
• Trained a transformer model (SyntMTE) on LM-generated reaction recipes and literature data.
• Evaluated model performance on a held-out dataset and a case study involving solid-state electrolytes (Li7La3Zr2O12).

Main Results:

• Off-the-shelf LMs achieved Top-1 precursor prediction accuracy of up to 53.8% and Top-5 of 66.8%.
• LMs predicted calcination and sintering temperatures with mean absolute errors <126 °C, outperforming specialized models.
• Ensembling LMs improved accuracy and reduced inference costs by up to 70%.
• A model trained solely on LM-generated data showed competitive performance, only 6% worse than literature-trained models.
• A hybrid model trained on both data types improved performance by up to 4% and reproduced experimental trends for solid-state electrolytes.

Conclusions:

• Language models possess significant, untapped potential for inorganic synthesis planning without task-specific fine-tuning.
• LM-generated data can augment limited literature datasets, leading to more robust and data-efficient training of synthesis prediction models.
• A hybrid approach combining LMs and transformer models offers a scalable and effective strategy for advancing inorganic synthesis planning.