Large Language Model-Guided Prediction Toward Quantum Materials Synthesis

TL;DR

A framework using large language models to predict synthesis pathways for inorganic materials, including quantum materials, demonstrates comparable performance across materials with varying degrees of quantumness quantified using quantum weight, indicating that LLMs offer a powerful tool to predict balanced chemical equations for quantum materials discovery.

Abstract

The synthesis of inorganic crystalline materials is essential for modern technology, especially in quantum materials development. However, designing efficient synthesis workflows remains a significant challenge due to the precise experimental conditions and extensive trial and error. Here, we present a framework using large language models (LLMs) to predict synthesis pathways for inorganic materials, including quantum materials. Our framework contains three models: LHS2RHS, predicting products from reactants; RHS2LHS, predicting reactants from products; and TGT2CEQ, generating full chemical equations for target compounds. Fine-tuned on a text-mined synthesis database, our model raises accuracy from under 40% with pretrained models, to under 80% using conventional fine-tuning, and further to around 90% with our proposed generalized Tanimoto similarity, while maintaining robust to additional synthesis steps. Our model further demonstrates comparable performance across materials with varying degrees of quantumness quantified using quantum weight, indicating that LLMs offer a powerful tool to predict balanced chemical equations for quantum materials discovery.

Figures (6)

• Figure 1: Overview of large language model prediction of the synthesis process. Synthesis protocols are pre-processed into structured data kononova2019text (the lower left figure is taken from the same reference), which specify target compounds, precursors, operations, and conditions. The synthesis pathway of BaTiO$_3$, a ferroelectric material, is shown as an example. The data is used to fine-tune large language models with three different models: (1) Left-hand-side-to-right-hand-side (LHS2RHS), where the model predicts products ($\rm{C + D}$) from given reactants ($\rm{A + B}$); (2) Right-hand-side-to-left-hand-side (RHS2LHS), predicting reactants from known products; and (3) Target-to-chemical-equation (TGT2CEQ), generating the full chemical equation from solely the target compound C. In the chemical equations, the black colors denote the input in each model, while the blue colors denote the output.
• Figure 2: Workflow to compute the generalized Tanimoto similarity (GTS). For any pair of predicted and ground truth chemical equations, each chemical formula, represented by A, B, ... J, from both equations is vectorized into vector of element counts $\mathbf{v}_\text{A},\mathbf{v}_\text{B},...\mathbf{v}_\text{J}$ and partitioned into reactants (yellow), and products (blue) sets. We calculate the Tanimoto similarity between every ordered pair of vectors from Equations 1 and 2, resulting in similarity matrices for reactants and products. Then, we extract Tanimoto set similarities ($T_s(\mathbf{R}_1,\mathbf{R}_2)$, $T_s(\mathbf{P}_1,\mathbf{P}_2)$) by iterating through rows of both matrices, selecting the maximum Tanimoto similarity for each row and averaging them ($T_s(\mathbf{R}_1|\mathbf{R}_2)$, $T_s(\mathbf{P}_1|\mathbf{P}_2)$). This effectively pairs each chemical formula from Equation 1 to a chemical formula from Equation 2 that is most similar. We then symmetrize the similarities by applying the same procedure as above, switching the roles of Equation 1 and Equation 2 ($T_s(\mathbf{R}_2|\mathbf{R}_1)$, $T_s(\mathbf{P}_2|\mathbf{P}_1)$). Finally, the GTS $\bar{T}(\mathbf{eq}_1, \mathbf{eq}_2)$ is calculated by averaging the Tanimoto set similarities of reactants and products.
• Figure 3: Schematics and average prediction accuracies for LHS2RHS and RHS2LHS. (a) LHS2RHS schematic, where reactants are given as input and products are generated. (b) RHS2LHS schematic, where products are given as input and reactants are generated. (c-d) GTS and JS for LHS2RHS predictions. (e-f) GTS and JS for RHS2LHS predictions. Models with (w/) fine-tuning (blue) perform significantly better than models (pink) without (w/o) fine-tuning. Our proposed GTS shows a notably higher accuracy than conventional JS accuracy.
• Figure 4: Schematics and average prediction accuracies for TGT2CEQ. (a) TGT2CEQ schematic, where the target compound is given as input and the full chemical equation is generated. (b-c) GTS and JS for models (blue) with (w/) fine-tuning and models (pink) without (w/o) fine-tuning. Our proposed GTS shows a consistent higher accuracy than using conventional JS.
• Figure 5: Robustness of chemical equation predictions with additional synthesis operation prompts. (a-c) Representations of LHSOPE2RHS, RHSOPE2LHS, and TGTOPE2CEQ where synthesis operations are included as additional prompts. (d-f) Prediction accuracy without (blue) and with (orange) these additional prompts using GTS. The preservation of prediction accuracy shows robustness against additional operation prompts.