StructLLM

This repository contains the data and code for "Explainable Synthesizability Prediction of Inorganic Crystal Polymorphs using Large Language Models"

Developer

Seongmin Kim, Joshua Schrier, and Yousung Jung

Organization

First, download the related data from zenodo (link: https://zenodo.org/records/14729225 ) and unzip.

data folder : general inorganic material data

perovskite folder : perovskite material data

result folder : model results for general synthesizability prediction and explanation

batch_request folder : batch data and batch result for OpenAI batch api request

Follow the codes in the order indicated by the numbered indexing. (From 00_Data.py to 06_get_metrics_perovskite.py)

Instructions

Python code (.py) uses python 3.11.8 and requires libraries;

Numpy (version == 1.26.4), Pymatgen (version == 2024.3.1), and OpenAI (version == 1.30.1).

For generating structural text description, Robocrystallographer (version == 0.2.8) was used.

You can manually install or using requirements.txt to use following codes:

• 00_Data.py : Data preprocessing of General inorganic material for training and finetuning.
• 01_predict_StructSynthGPT-FT_batch.py : Using fine-tuned StructLLM, predict synthesizability of hold-out-test dataset by OpenAI batch api request.
• 01a_batchresult2resultformat.py : Convert the batch result ('01_predict_StructSynthGPT-FT_batch.py') to data analysis format.
• 02_get_metrics.py : See the result of models for general synthesizability prediction
• 03_explain_Synthesizability_GPT4o_4om.py : Using GPT-4o, explain the reasons of general synthesizability by OpenAI batch api request.
• 03a_batchresult2resultformat_explanation_4om.py : Convert the batch result ('03_explain_Synthesizability_GPT4o_4om.py') to data analysis format.
• 04_get_explanation_GPT_embedding_4om.py : Using gpt text-embedding-large model, get embedding vectors of textual explanations ('03a_batchresult2resultformat_explanation_4om.py').
• 04a_make_GPTembedding_json_dictionary_4om.py : Convert the batch result ('04_get_explanation_GPT_embedding_4om.py') to data analysis format.
• 04b_explanation_analysis_4om.py : See the explanation result of models for general synthesizability prediction
• 05_predict_PerovStructGPT-FT-TL_batch.py : Using fine-tuned perovskite StructLLM, predict synthesizability of hold-out-test perovskite dataset by OpenAI batch api request.
• 06_get_metrics_perovskite.py : See the result of models for perovskite synthesizability prediction

Code guideline

Follow the below steps to use the code and to reproduce this method as it is;

• Get your OpenAI api key. (See https://platform.openai.com/)
• Download MP30_description dataset.
• Fix config.json file. ( Type your OpenAI api key & your downloaded mp30_description folder path )
• Run 00_Data.py code. --> ./data/finetuning/train_pu_struct_data.jsonl, ./data/finetuning/val_pu_struct_data.jsonl, ./data/finetuning/hold_out_pu_struct_data.jsonl will be generated.
• Fine-tuning GPT-4o-mini by using the generated ./data/finetuning/train_pu_struct_data.jsonl, ./data/finetuning/val_pu_struct_data.jsonl dataset. (See https://platform.openai.com/docs/guides/fine-tuning/)
• After finishing the fine-tuning, copy your fine-tuned model identifier and paste it into 01_predict_StructSynthGPT-FT_batch.py requesting model argument.
• Run 01_predict_StructSynthGPT-FT_batch.py code. --> Batch request will be calculated. After batch completed, you can download the output jsonl file.
• Copy your batch identifier and paste it into 01a_batchresult2resultformat.py at batch output load part.
• Run 01a_batchresult2resultformat.py to convert batch output to readable json file. --> The test result will be saved as ./result/prediction_StructSynthGPT4om.json.

If you want to test another inorganic structures, make the same format to the test dataset (./data/finetuning/hold_out_pu_struct_data.jsonl) and run 01_predict_StructSynthGPT-FT_batch.py code. To obtain the structural text from cif file, use Robocrystallographer tool. (See https://hackingmaterials.lbl.gov/robocrystallographer/index.html#)

Cite

A publication appears on the Angewandte Chemie International Edition as doi:10.1002/anie.202423950