Skip to content

PaddleMaterials_MetaX_README.md

Latest commit

 

History

History
116 lines (101 loc) · 6.03 KB

PaddleMaterials_MetaX_README.md

File metadata and controls

116 lines (101 loc) · 6.03 KB

PaddleMaterials on GiteeAI (MetaX)

Environment Setup on GiteeAI

  1. Register and log in to giteeAI.
  2. Purchase computing resources and click Rent Now.
  3. Choose the PaddleMaterials image and click Next to create an instance.
  4. After the instance is created, click Lab to enter the container.
  5. In the Lab page, choose Jupyter Lab, then open a Terminal.

Training Process

PaddleMaterials source directory: /opt/PaddleMaterials Reference documents:

Below is the Structure Generation / DiffCSP example.

1) Train

# single-gpu training
python structure_generation/train.py -c structure_generation/configs/diffcsp/diffcsp_mp20.yaml

2) Sample

python structure_generation/sample.py --model_name='diffcsp_mp20' --weights_name='latest.pdparams' --save_path='result_diffcsp_mp20-1/' --chemical_formula='LiMnO2'

Result example: Li1-Mn1-O2_1.cif

MetaX Results (from the "PaddleMaterials on MetaX benchmark" sheet)

Machine Learning Interatomic Potentials (MLIP)

Model Metric Benchmark MetaX Difference Notes
CHGNet energy_per_atom -7.367691 -7.3676915 5.0e-07
CHGNet force (max abs diff) - - 3.03e-06 Max component diff across 8 samples; see sheet for per-axis values
CHGNet magmom (max abs diff) - - 7.0e-07 Max across 8 samples; see sheet for per-axis values
MatterSim energy -6.6172876 -6.6172876 0
MatterSim force (max abs diff) - - 1.64e-06 Max component diff across 8 samples; see sheet for per-axis values

CHGNet - force details

# Benchmark (x, y, z) MetaX (x, y, z) Difference (x, y, z)
1 -2.98e-08, -1.13e-07, 2.38e-02 1.19e-07, -7.66e-08, 2.38e-02 1.49e-07, 3.62052e-08, 8.214e-07
2 2.98e-08, -1.41e-07, -2.38e-02 -4.47e-08, 7.60e-08, -2.38e-02 7.45058e-08, 2.16824e-07, 5.809e-07
3 3.58e-07, 1.12e-07, 9.26e-02 1.49e-07, 4.66e-08, 9.26e-02 2.08616e-07, 6.51926e-08, 3.879e-07
4 -2.98e-08, 4.10e-08, -9.26e-02 1.49e-07, -1.49e-08, -9.26e-02 1.78814e-07, 5.58794e-08, 3.75e-08
5 8.94e-08, 1.46e-07, -2.43e-03 2.98e-08, 6.89e-08, -2.43e-03 5.96046e-08, 7.68341e-08, 2.3842e-07
6 -2.09e-07, -1.97e-06, -1.31e-02 -7.45e-07, 1.06e-06, -1.31e-02 5.36442e-07, 3.0254e-06, 1.5495e-06
7 2.98e-08, 1.89e-06, 1.31e-02 6.26e-07, -1.03e-06, 1.31e-02 5.96046e-07, 2.91504e-06, 1.0731e-06
8 -1.19e-07, 7.73e-08, 2.43e-03 -5.96e-08, -6.19e-08, 2.43e-03 5.96046e-08, 1.39233e-07, 1.22192e-06

CHGNet - magmom details

# Benchmark MetaX Difference
1 3.04922e-03 3.05e-03 6.71e-08
2 3.04934e-03 3.05e-03 1.043e-07
3 3.86942e+00 3.87e+00 7e-07
4 3.86942e+00 3.87e+00 7e-07
5 4.41358e-02 4.41e-02 1.64e-07
6 3.86221e-02 3.86e-02 0
7 3.86220e-02 3.86e-02 1.26e-07
8 4.41357e-02 4.41e-02 7.4e-08

MatterSim - force details

# Benchmark (x, y, z) MetaX (x, y, z) Difference (x, y, z)
1 1.72e-07, -9.79e-08, 1.26e-01 7.28e-08, 3.45e-08, 1.26e-01 9.8953e-08, 1.32313e-07, 9e-08
2 7.01e-08, 5.59e-09, -1.26e-01 -1.08e-07, 8.89e-08, -1.26e-01 1.77938e-07, 8.3357e-08, 7.4e-07
3 -8.96e-09, -1.93e-07, -1.51e-01 -1.87e-08, -6.15e-07, -1.51e-01 9.77889e-09, 4.22704e-07, 3.1e-07
4 -2.14e-07, -1.66e-07, 1.51e-01 -1.33e-07, 2.61e-07, 1.51e-01 8.07632e-08, 4.26546e-07, 1.01e-06
5 1.65e-07, 2.76e-07, -1.01e-01 -5.97e-08, 3.58e-07, -1.01e-01 2.25031e-07, 8.19564e-08, 1.4e-06
6 -1.81e-07, 2.43e-07, 8.90e-02 -3.03e-07, -9.78e-08, 8.90e-02 1.21773e-07, 3.41037e-07, 3.57e-07
7 -1.89e-07, 4.66e-08, -8.90e-02 5.16e-07, 1.99e-07, -8.90e-02 7.04895e-07, 1.52737e-07, 5.66e-07
8 -1.68e-07, -1.27e-07, 1.01e-01 -3.02e-08, -2.35e-07, 1.01e-01 1.38127e-07, 1.08732e-07, 1.64e-06

Property Prediction (PP)

Model Metric Benchmark MetaX Difference
MEGNet formation_energy_per_atom -2.1585608 -2.1585603 5.0e-07
MEGNet band_gap 1.11667 1.1166712 1.0e-07
MEGNet shear_modulus 4.02312e+02 402.3117952 0
MEGNet bulk_modules 1.10824e+03 1108.240846 0
ComFormer formation_energy_per_atom -2.1615877 -2.1615882 5.0e-07
ComFormer band_gap 1.06078 1.0607839 3.0e-07
ComFormer shear_modulus 459.992749 4.59993e+02 0
ComFormer bulk_modules 1.02844e+03 1028.443467 0
DimeNet++ formation_energy_per_atom -2.1553288 -2.1553288 0
DimeNet++ band_gap 1.03222 1.0322216 4.0e-07
DimeNet++ shear_modulus 6.77118e+01 67.71178 2.0e-05
DimeNet++ bulk_modules 1.06495e+02 106.4948 3.0e-05

Structure Generation (SG)

Model Metric Benchmark MetaX Difference
MatterGen S.U.N. Functional Functional -
DiffCSP match_rate 5.16914e-01 0.514370993 0.00254256
DiffCSP rms_dist 5.77073e-02 0.058631274 0.000924018

Spectrum Elucidation (SE)

Model Metric Benchmark MetaX Difference
DiffNMR Rebuild accuracy 5.06670e-01 0.50535 0.00132