Bridging the Semantic-Numerical Gap: A Numerical Reasoning Method of Cross-modal Knowledge Graph for Material Property Prediction

Official code for "Bridging the Semantic-Numerical Gap: A Numerical Reasoning Method of Cross-modal Knowledge Graph for Material Property Prediction"

Dependencies

conda create -n NRKG python=3.8
conda activate NRKG

# Please install PyTorch according to your CUDA version.
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 cpuonly -c pytorch

pip install -r requirements.txt

HEA dataset

We proposed two high-entropy alloy (HEA) datasets, HEA-HD and HEA-CRD. HEA-HD is a hardness dataset of HEAs, and HEA-CRD is a corrosion resistance dataset of HEAs.

We also provide the cross-modal knowledge graph based on the dataset, which is in ./dataset/HEA-HD-KG and ./dataset/HEA-CRD-KG. The entity.pkl and relation.pkl are the entity and relation information. Please use the pickle library to read.

NR-KG Framework

Project Directory Structure

NR-KG
├───dataset
│   ├───HEA-CRD-KG
│   ├───HEA-CRD-Ori
│   ├───HEA-HD-KG
│   └───HEA-HD-Ori
├───model
│   └───pre_trained
│       ├───HEA-CRD
│       └───HEA-HD
├───res
├───tools
├───utils
├───GCN_model.py
├───main.py
├───requirements.txt
├───LICENSE
└───README.md

Train NR-KG

To train the NR-KG model on the HEA-HD dataset, run the following command:

python main.py --root_path ./ --dataaddr ./dataset --n_fold 6 
--seed 523 --data_seed 2021 --id 0 --epoch 2000 --fp16 False --hidden_size 128 --num_layers 2 --dropout 0.5 --activation LeakyReLU --lr 0.005 --scheduler StepLR --gamma 0.997 --step_size 1 --a_CLL_loss 0.16 --a_MSELoss 1.0 --a_PPL_Loss 0.04 --patience 50 --early_stop True --KGtype HEA-HD-KG --gpu 0 --state train

To train the NR-KG model on the HEA-CRD dataset, run the following command:

python main.py --root_path ./ --dataaddr ./dataset  --n_fold 6 --seed 23 --data_seed 2024 --id 1 --epoch 2000 --fp16 False --hidden_size 128 --num_layers 2 --dropout 0.5 --activation LeakyReLU --lr 0.001 --scheduler StepLR --gamma 0.997 --step_size 1 --a_CLL_loss 0.16 --a_MSELoss 1.0 --a_PPL_Loss 0.02  --patience 50 --early_stop True --KGtype HEA-CRD-KG --gpu 0 --state train

Test NR-KG

To test the NR-KG model on the HEA-HD dataset using the checkpoint, run the following command:

python main.py --root_path ./ --dataaddr ./dataset --net_path ./model/pre_trained/HEA-HD --n_fold 6 --data_seed 2021 --id 5005 --KGtype HEA-HD-KG --gpu 0 --state test

To test the NR-KG model on the HEA-CRD dataset using the checkpoint, run the following command:

python main.py --root_path ./ --dataaddr ./dataset  --net_path ./model/pre_trained/HEA-CRD --n_fold 6 --data_seed 2024 --id 5005 --KGtype HEA-CRD-KG --gpu 0 --state test

The test results will be saved in the ./res folder. The results in this paper are run on NVIDIA RTX 1080Ti.

HEA-HD Dataset Results: The results of 6-fold cross-validation are as follows:

Fold	MAE	RMSE	R2
0	2958	38.72	0.92
1	4467	47.27	0.85
2	2677	37.93	0.90
3	3592	44.83	0.88
4	4793	48.13	0.84
5	2636	40.90	0.89
Ave. ± Std.	3520±931	42.96±4.39	0.88±0.03

HEA-CRD Dataset Results: The results of 6-fold cross-validation are as follows:

Fold	MAE	RMSE	R2
0	2.55	1.16	0.43
1	2.14	1.19	0.66
2	1.73	0.99	0.58
3	1.52	1.01	0.46
4	1.38	0.98	0.73
5	3.94	1.40	0.33
Ave. ± Std.	2.21±0.95	1.12±0.16	0.53±0.15

Citation

If you find this code useful in your research, please consider citing our paper.

bib format will be provided after the paper is published.

Contact

If you have any questions, feel free to contact me via issue or email (sguangxuan@163.com).