Question about input parameters for the E3 case

[choi-geunseok]

Details

What are the SLEM parameters for the Si case in the ICLR paper?
I am also curious about the settings for other materials like GaN and HfO₂.
Currently, the Hamiltonian error is at the level of 1e-2 eV, so I am not sure exactly what should be improved.

[floatingCatty]

Hello!

Thanks for considering our package. Here is the parameter used to train the GaN:

{
"common_options": {
"seed": 12342,
"basis": {
"Ga": "2s2p2d1f",
"N": "2s2p1d"
},
"device": "cuda",
"dtype": "float32",
"overlap": false
},
"model_options": {
"embedding": {
"method": "slem",
"r_max": {"Ga":8.1, "N":7.1},
"irreps_hidden": "64x0e+32x1o+16x2e+8x3o+8x4e+8x5o+4x6e",
"n_layers": 4,
"n_radial_basis": 18,
"env_embed_multiplicity":10,
"avg_num_neighbors": 51,
"latent_dim": 64,
"latent_channels": [32, 32],
"tp_radial_emb": true,
"tp_radial_channels": [32]
},
"prediction":{
"method": "e3tb",
"scales_trainable":false,
"shifts_trainable":false
}
},
"train_options": {
"num_epoch": 16000,
"batch_size": 1,
"optimizer": {
"lr": 0.005,
"type": "Adam"
},
"lr_scheduler": {
"type": "rop",
"factor": 0.8,
"patience": 50,
"min_lr": 1e-6
},
"loss_options":{
"train":{"method": "hamil_abs"},
"validation":{"method": "hamil_abs"}
},
"save_freq": 1000,
"validation_freq": 10,
"display_freq": 1
},
"data_options": {
"train": {
"type": "DefaultDataset",
"root": "/personal/Hamil_datasets/GaN-300/cubic_300K/train/",
"prefix": "frame",
"get_Hamiltonian": true,
}
}
}

The codebase has been updated with several large refactorings and method improvements since then; it might have discrepancies, but the accuracy should be better than what is reported in the paper.