MatAltMag

The pytorch implement of papar AI-accelerated Discovery of Altermagnetic Materials

Prerequisites

package required:

• pytorch
• accelerate
• pymatgen
• PyYAML
• tensorboard
• tqdm
• pandas

Run

1. run files in the preprocess directory, move the output files of label0.csv and candidate.csv to root_dir directory
2. check required data files in root_dir

• atom_init.json: a JSON file that stores the initialization vector for each element.
• label0.csv: a CSV file that stores the ID for the non-altermagnetic crystal
• label1.csv: a CSV file that stores the ID for the altermagnetic crystal
• candidate.csv: a CSV file that stores the ID for the crystal in candidate datasets

3. run download.py to download CIF files of all crystals in the three CSV files from Materials Project. Download time depends on your internet speed. Once completed, the structure under root_dir will be

root_dir
├── atom_init.json
├── label0.csv
├── label1.csv
├── candidate.csv
├── id_prop_0.csv
├── id_prop_1.csv
├── id_prop_-1.csv
├── id0.cif
├── id1.cif
├── ...

4. set the configuration of accelerate in a proper path, for example

accelerate config --config_file yamls/accelerate.yaml

5. check and update yamls/pretrain.yaml, then run pretrain.py

nohup sh pretrain.sh &

or

accelerate launch --config_file yamls/accelerate.yaml pretrain.py --file yamls/pretrain.yaml 

6. check and update yamls/train.yaml, then run train.py

nohup sh train.sh &

or

accelerate launch --config_file yamls/accelerate.yaml train.py --file yamls/train.yaml

7. check and update yamls/predict.yaml, then run predict.py

python predict.py --file yamls/predict.yaml

You can also directly load our trained model without pre-training and training yourself. Our weights of classifier model can be download from Google Drive.