Low-energy Latent Space Search Workflow

Tasks

Tasks in the LOLS workflow.

• main
  • data_generation
    • train_vae
    • sample
    • dft_energy
  • energy_model
    • gp_model
    • gp_minimum
  • relaxation
    • dft_relax

Read our publication Molecular Conformer Search with Low-Energy Latent Space for details.

Run

1. Switch to the python env which have pytorch and BOSS installed.

• pytorch version: 1.3.1
• BOSS version: 0.9.17

2. Create config.json.

• Check the template files in lib/config.
• If there's no config.json, the config-test.json will be used.

3. Run python main.py.

Restart Task

Set task's state = 0 or 1 to restart that task:

sqlite3 lols.db 'update task set state = {new_state} where id = {task_id}'

• If the state set to 0, the task will refresh its config before running.

Database

table task:

column	information
id	task id, the primary key
state	0: not start, 1: start, 2: finish
iteration	number of iteration
key	sub index
name	name of task, related to task/{name}.py
config	config of task, will save in config.json

Folders and Files

folder	information
lib	read-only files for materials and software configure
model	Encapsulation of VAE model and other physical models.
task	LOLS tasks
work	Isolated folders for all tasks.

file	information
config.json	Where configurations for all tasks are saved.
lols.db	data base, where only the task information are saved.
database.py	python modules for operating lols.db
main.py	entries for the program

Hack BOSS

We need only the Gaussian Process fitting hyper_parameters, not the x_next and global_minimum. So we hacked BOSS library to add the s model, by replaced the source code with lib/boss_modify/*.

Since the current boss version is 1.5 (Checked at 2022-06-24), it might provide the options to skip finding the x_next or global_minimum.

Use it at your own risk.

You can also replace lib/gp_model.py to use the GPy library directly or other Gaussian Process library like gpytorch.