Reinforcement Learning Framework For MOFs

This package is a reinforcement learning framework for MOFs. The framework consists of agent and environment which are a generator and a predictor, respectively. The agent takes an action (which is generating a MOF structure). This action is then evaluated in the environment by the predictor, which predicts the value of the property we are interested in. Based on the prediction, a reward is returned in form of an update to the agent to generate the next round of MOFs.

Installation

OS and hardware requirements

Linux : Ubuntu 20.04, 22.04

For optimal performance, we recommend running with GPUs

Dependencies

python>=3.8

Install

Please install pytorch (>= 1.12.0) according to your environments before installation of requirements.

pip install -e .

Getting Started

download pre-trained models

In order to run the reinforcement learning framework, predictor (environment) and generator (agent) should be pre-trained.

mofreinforce download default

Then, you can find the pre-trained generator and predictors in mofreinforce/model.

Predictor

If you want to train the predictor for your own desired property, please refer to predictor.md.

Generator

We provide a generator which selects a topology and a metal cluster, which are categorical variables, in order and then creates an organic linker represented by SELFIES string. The generator was pre-trained with about 650,000 MOFs created by PORMAKE, which allows for generating feasible MOFs. You can find the pre-trained generator at model/generator.ckpt.

Reinforcement Learning

(1) reinforcement learning with CO2 heat of adsorption

python run.py with v0_qkh

(2) reinforcement learning with CO2/H2O selectivity

python run.py with v1_selectivity

if you want to experiment with other parameters by modifying mofreinforce/reinforce/config_freinforce.py