Step 0: Setup: LMDBs, trajectories, and metadata files should be downloaded.
Step 1: Run ML relaxations on the LMDB to generate predictions of relaxed energy and relaxed structures. For example, see these instructions. We will be using the trajectory files.
Step 2: Process ML relaxed structures (MLRS), group by system, and sort configurations by energy.
python process_mlrs.py \
--ml-trajs-path $MODEL_ROOT_DIR/ml_trajs/ \
--outdir $MODEL_ROOT_DIR/cache/ \
--metadata $DATA_ROOT_DIR/mappings/oc20dense_mapping.pkl \
--surface-dir $DATA_ROOT_DIR/trajs/
Step 3: Select the best k and write DFT input files (VASP used in this work) There are different parameters for single points vs. relaxations.
SP="--sp --nsw 0 --nelm 300"
RX="--nsw 2000 --nelm 60"
python write_top_k_vasp.py \
--cache $MODEL_ROOT_DIR/cache/cache_sorted_byE.pkl \
--outdir $MODEL_ROOT_DIR/dft/ \
--k 5 $SP
Step 4: Run DFT single points or relaxations on best k.
Be sure to filter for convergence, see utils.py for example code.
Step 5: Organize ML and DFT results in the required format outlined here.
A sample submission file can be downloaded here: https://dl.fbaipublicfiles.com/opencatalystproject/data/adsorbml/gemnet_oc_ml_sp_sample_results.tar. Note - configurations failing the physical constraints were excluded from the submission file, do NOT rely on matching each entry. This is intended to provide you with an idea to the format.
Step 6: Evaluate results with dense_eval.py with the specified format of predictions to get the success rate and DFT speedup.