Learning Superconductivity from Ordered and Disordered Material Structures (NeuIPS 2024)

Superconductors Inverse Design Pipeline

[Paper]

Step 1: Prepare the dataset

Prepare the raw data as

|-- data
    |-- properties
	    |-- <property>
            |-- cif
            |-- <raw_data>.csv

The csv file should at least contain the following 3 columns

material_id, cif, <prop>

<prop> can be arbitrary property types, like Tc in superconductors.

Split the raw data via the following script

python scripts/make_split.py --dir data/properties/<property> --csv <raw_data>.csv

The default setting will shuffle the dataset in random seed 42 and split it into train.csv, val.csv and test.csv with ratio 8:1:1.

Step 2: Train a property prediction model

python diffcsp/run.py model=prediction data=property data.subdir=<property> data.prop=<prop> data.task=<task> data.opt_target=<opt_target> exptag=<property>_<prop> expname=prediction

The trained model is saved in singlerun/<property>_<prop>/prediction. The default 3D encoder is DimeNet++, and one can change it into more powerful encoders (e.g. Equiformer).

<task> can be chosen from classification/regression.

<opt_target> have different meanings for different tasks:

For classification, <opt_target> means the required class to generation. For regression, <opt_target> = 1 means to generate candidates with higher property (like Tc), while <opt_target> = -1 means to generate candidates with lower property (like formation energy)

Step 3: Train a time-dependent guidance model

python diffcsp/run.py model=guidance data=property data.subdir=<property> data.prop=<prop> data.task=<task> data.opt_target=<opt_target> exptag=<property>_<prop> expname=guidance

The trained model is saved in singlerun/<property>_<prop>/guidance.

Step 4: Generate candidates with guidance

python scripts/optimization.py --model_path ${PWD}/singlerun/<property>_<prop>/guidance --uncond_path ${PWD}/singlerun/2023-04-18/pure_pretrain

The above command will yield eval_opt.pt under the singlerun/<property>_<prop>/guidance directory, which contains 500 optimized structures.

Step 5: Evaluate the trained model and optimized samples

python scripts/eval_optimization.py --dir ${PWD}/singlerun/<property>_<prop>

The results are logged in singlerun/<property>_<prop>/results as

|-- results
    |-- summary.log
    |-- results.csv
    |-- cif
        |-- xx.cif
        ...

summary.log summaries the results of the property prediction & guidance model. An example is provided as

*************** Property Prediction ***************

Test pcc: 0.4857

*************** Optimization ***************

Top-5 Results: 
489-xx: xx
385-xx: xx
249-xx: xx
486-xx: xx
163-xx: xx

An example script for the entire pipeline

export CUDA_VISIBLE_DEVICES=1

python scripts/make_split.py --dir data/properties/SuperCon --csv order_data_tc.csv

python diffcsp/run.py model=prediction data=property data.subdir=SuperCon data.prop=logtc data.task=regression data.opt_target=1 exptag=SuperCon_logtc expname=prediction

python diffcsp/run.py model=guidance data=property data.subdir=SuperCon data.prop=logtc data.task=regression data.opt_target=1 exptag=SuperCon_logtc expname=guidance

python scripts/optimization.py --model_path ${PWD}/singlerun/SuperCon_logtc/guidance --uncond_path ${PWD}/singlerun/2023-04-18/pure_pretrain

python scripts/eval_optimization.py --dir ${PWD}/singlerun/SuperCon_logtc

Citation

Please consider citing our work if you find it helpful:

@inproceedings{chenlearning,
  title={Learning Superconductivity from Ordered and Disordered Material Structures},
  author={Chen, Pin and Peng, Luoxuan and Jiao, Rui and Mo, Qing and Zhen, WANG and Huang, Wenbing and Liu, Yang and Lu, Yutong},
  booktitle={The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2024}
}