MatAltMag

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

Prerequisites

some package required:

• pytorch: 2.0.1
• accelerate: 0.20.0
• pymatgen: 2023.5.10
• PyYAML: 6.0
• tqdm: 4.64.0

Other required packages are listed in the requirements.txt file.

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

Downloading all CIF files of all crystals using the download.py script takes about 1 hour, depending on your network speed. Pretraining the auto-encoder model for 10 epochs with a batch size of 64 on 2 NVIDIA A100 GPUs takes over 2 days. Predicting the candidate datasets of over 42,000 samples takes about 10 seconds with a batch size of 512.

You can also directly load our trained model, which has undergone multiple iterative training processes, without pre-training and training it yourself. The weights of our classifier model can be downloaded from Google Drive. The corresponding output is presented in the out/output.csv. In addition, over three hundred additional candidate materials (unconfirmed yet by DFT calculations) were predicted by the proposed AI search engine. These candidates were listed in out/Candidate_for_DFT_validate.csv.