MPF.2021.2.8

[Reverting back to v1, please make sure to use an earlier pymatgen version.] <br> This dataset contains the MPF.2021.2.8 data used to train the m3gnet model reported in `https://arxiv.org/abs/2202.02450` <br> I have split the dataset into two pickle files. To load the data, you can use example code as below. <br> ``` import pickle with open('block_0.p', 'rb') as f: data = pickle.load(f) <br> with open('block_1.p', 'rb') as f: data.update(pickle.load(f)) ``` <br> where `data` will be a dictionary with `material_id` as the key and an inner dictionary as the value. <br> The inner dictionary contains the snapshots of this `material_id`, with the following keys. ``` - structure - energy - force - stress - id ``` The `structure` is a list of pymatgen structures. <br> Each id in the `id` list is of format `material_id-calc_id-ionic_step_id`, where `calc_id` is 0 (second) or 1 (first) in the double relaxation process. <br> The `stress` here is the raw output from VASP, meaning that it is really the negative stress using the convention in our paper. Hence to train the model, please multiply stress with -0.1 (kBa to GPa and change sign) <br> The units for energy, force and stress in the data are eV, eV/A, and kBa. Remember to convert the stress to GPa and take the negative sign to work with m3gnet training.