README.md

metadata

tags:
  - chemistry

Description

VSCF MLFF is the code repository for the paper "Интерполяция ППЭ с помощью машинного обучения для ускорения расчётов негармонических частот колебаний молекул".

You can find here these files

• compare_models.py -- script for comparing SchNet and DimeNetPlusPLus models for PES predicting
• train.py -- script for training the model
• example_plot.py -- script, which makes example plots by statistics emitted by train.py file with _metrics.pkl suffix
• example_eval.py -- script, which evaluates model for the PES of water molecule, using model emittet by train.py script, and fills GAMESS-US input template with it.
• utils_data.py and utils_model.py -- libraries with functions commonly used by scripts from this repository

Usage

compare_models.py

python compare_models.py <filename> <model>

• filename trajectory file emitted by xtb program
• model SchNet or DimeNet

train.py

python train.py [--normalize] [--pretrain] [--loss_k <float>] <loss_mode> <checkpoint_prefix> <data_filename>

• normalize -- normalize the target (can only be used with loss_mode=mae)
• pretrain -- add pretraining stage in addition to the finetuing stage
• loss_k -- k coefficient for adaptive loss (can only be used for adaptive loss)
• loss_mode -- mae or adaptive
• checkpoint_prefix -- prefix for checkpoint files
• data_filename -- file name of the dataset of molecular VSCF grids

example_plot.py

python example_plot.py pickle_file

• pickle_file -- path to the *_metrics.pkl file

example_dara.py

python example_plot.py [--denormalize] <model_filename> <data_filename>

• denormalize -- denormalize the target (use if your model if trained with --normalize)
• model_filename -- path to the model checkpoint file
• data_filename -- file name of the dataset of molecular VSCF grids