T2BR protocol

Implementation of "Text-to-Battery Recipe: A language modeling-based protocol for automatic battery recipe extraction and retrieval"

Recent studies have increasingly applied natural language processing (NLP) to automatically extract experimental research data from the extensive battery materials literature. Despite the complex process involved in battery manufacturing — from material synthesis to cell assembly — there has been no comprehensive study systematically organizing this information. In response, we propose an NLP-based protocol, Text-to-Battery Recipe (T2BR), for the automatic extraction of end-to-end battery recipes, validated using a case study on batteries containing LiFePO4 cathode material.

Workflow of T2BR protocol

Protocol structure and dataset

Dataset and pre-trained models used in this study is available here, and unzip in your own directory.

Data preparation

• data/raw_paper.csv: the abstracts of total papers.
• data/doc_label.xlsx: the annoation for text classification.
• data/doc_topic.xlsx: the paragraph - topic allocation results.
• data/raw_paragraphs.xlsx: the total paragraphs of LFP battery papers.
• data/topic_kwd.xlsx: the main keywords of topics.

RecipeExtraction

• data/cathode_synthesis.json: the annotation dataset for cathode synthesis NER model training.
• data/cell_assembly.json: the annotation dataset for cell assembly NER model training.
• utils/phraser.pkl: MaterialsTextTokenizer requires this file.
• models/best_model/cathode_synthesis/best.pt: pre-trained NER model for cathode synthesis.
• models/best_model/cell_assembly/best.pt: pre-trained NER model for cell assembly.

RecipeSearch

• data/cathode_recipes.xlsx: the recipe sequences of cathode synthesis.
• data/cell_recipes.xlsx: the recipe sequences of cell assembly.
• data/full_recipes.xlsx: the end-to-end recipes of batteries.

Requirements

Our experiment setting is as follows:

• gensim : 4.3.2
• nltk : 3.8.1
• h20 : 3.46.0.1
• seqeval : 1.2.2
• transformers : 4.32.1
• torchcrf : 1.1.0
• torchtools : 0.1.4
• chemdataextractor : 1.3.0
• monty: 2023.9.5
• pymatgen: 2023.8.10

Install

pip install -r requirements.txt

PLM configuration

from transformers import *

#BERT from arXiv preprint arXiv:1810.04805.
tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")
model = AutoModelForMaskedLM.from_pretrained("google-bert/bert-base-uncased")

#SciBERT from arXiv preprint arXiv:1903.10676.
tokenizer = AutoTokenizer.from_pretrained('allenai/scibert_scivocab_uncased')
model = AutoModel.from_pretrained("allenai/scibert_scivocab_uncased")

#BatteryBERT from Journal of chemical information and modeling, 62(24), 6365-6377.
tokenizer = AutoTokenizer.from_pretrained("batterydata/batterybert-uncased")
model = AutoModelForMaskedLM.from_pretrained("batterydata/batterybert-uncased")

#MatBERT from Patterns 3.4 (2022).
#The pre-trained MatBERT model is publicly at https://figshare.com/articles/software/MatBERT-NER_models/15087276.
#Download and unzip this file for your directory.
tokenizer = BertTokenizer.from_pretrained('./matbert-base-uncased')
model = BertForMaskedLM.from_pretrained('./matbert-base-uncased')

#ULSA from Digital Discovery, 1(3), 313-324.
#Download from https://github.com/CederGroupHub/synthesis-action-retriever
from synthesis_action_retriever.synthesis_action_retriever import SynthActionRetriever
dir_path = "path-to-models"
w2v_model = "path-to-w2v_model"
ext_model = "path-to-ext_model"
sar = SynthActionRetriever(
    embedding_model=os.path.join(dir_path, w2v_model),
    extractor_model=os.path.join(dir_path, ext_model)
)
with open('./data/example_sentences.json', 'r') as fp:
    examples = json.load(fp)
for sent in examples:
    actions = sar.get_action_labels(spacy_tokens)  

RecipeExtraction

NER model train

python train.py --device gpu:0 --seeds 1,2,3 --tag_schemes iobes --splits 80 --datasets cell_assembly --models matbert --batch_size 5 --optimizer_name rangerlars --weight_decay 0.001 --n_epoch 50 --embedding_unfreeze 1 --transformer_unfreeze 0,12 --embedding_learning_rate 1e-4 --transformer_learning_rate 2e-3 --classifier_learning_rate 1e-2 --scheduling_function exponential --keep_model

NER model test

python extract.py --model batterybert --type cell_assembly --save_dir ./result

RecipeSearch

python search.py -query <query> -recipe_type <type>

query: "(('sucrose'). PREC.) AND (('solid state method'). METHOD)"
recipe_type: end
There are 1 recipes in our end database.

        step_1: {'METHOD': ['solid state method'], 'TARGET_MATERIAL': ['LiFePO4'], 'act_type': ['Starting'], 'type': 'cathode'}
        step_2: {'ATMOSPHERE': ['H2'], 'COMPANY': ['aldrich'], 'PRECURSOR': ['FeC2O42H2O', 'LiH2PO4', 'citric acid', 'sucrose'], 'SOLVENT': ['acetone'], 'SPEED': ['400 rpm'], 'TIME': ['6 h'], 'act_type': ['Mixing', 'Mixing', 'Mixing', 'Mixing'], 'type': 'cathode'}

        ...

        step_20: {'ANODE': ['Li'], 'act_type': [], 'type': 'cell'}
        step_21: {'AMOUNT': ['1 m'], 'act_type': [], 'type': 'cell'}

Citation

If you find the codes and dataset useful, please cite our paper:

title = "Text-to-Battery Recipe: A language modeling-based protocol for automatic battery recipe extraction and retrieval",
authors = Daeun Lee, Jaewoong Choi, Hiroshi Mizuseki, Byungju Lee
journal = "",
year = "2024",
doi = ""