PMTransformer (MOFTransformer)

This package provides a universal transfer learning model, PMTransformer (Porous Materials Transformer), which obtains the state-of-the-art performance in predicting various properties of porous materials. The PMTRansformer was pre-trainied with 1.9 million hypothetical porous materials including Metal-Organic Frameworks (MOFs), Covalent-Organic Frameworks (COFs), Porous Polymer Networks (PPNs), and zeolites. By fine-tuning the pre-trained PMTransformer, you can easily obtain machine learning models to accurately predict various properties of porous materials .

NOTE: From version 2.0.0, the default pre-training model has been changed from MOFTransformer to PMTransformer, which was pre-trained with a larger dataset, containing other porous materials as well as MOFs. The PMTransformer outperforms the MOFTransformer in predicting various properties of porous materials.

Release Note

Version: 2.2.0 Now, MOFTransformer support multi-task learning (see multi-task learning)

Install

Depedencies

NOTE: This package is primarily tested on Linux. We strongly recommend using Linux for the installation.

python>=3.8

Given that MOFTransformer is based on pytorch, please install pytorch (>= 1.12.0) according to your environments.

Installation using PIP

$ pip install moftransformer

Download the pretrained models (ckpt file)

you can download the pretrained models (PMTransformer.ckpt and MOFTransformer.ckpt) via figshare

or you can download with a command line:

$ moftransformer download pretrain_model

(Optional) Download pre-embeddings for CoREMOF, QMOF

we've provide the pre-embeddings (i.e., atom-based graph embeddings and energy-grid embeddings), inputs of PMTransformer, for CoREMOF, QMOF database.

$ moftransformer download coremof
$ moftransformer download qmof

Getting Started

Install GRIDAY to calculate energy-grids from cif files

$ moftransformer install-griday

Run prepare-data .

from moftransformer.examples import example_path
from moftransformer.utils import prepare_data

# Get example path
root_cifs = example_path['root_cif']
root_dataset = example_path['root_dataset']
downstream = example_path['downstream']

train_fraction = 0.8  # default value
test_fraction = 0.1   # default value

# Run prepare data
prepare_data(root_cifs, root_dataset, downstream=downstream, 
             train_fraction=train_fraction, test_fraction=test_fraction)

Fine-tune the pretrained MOFTransformer.

import moftransformer
from moftransformer.examples import example_path

# data root and downstream from example
root_dataset = example_path['root_dataset']
downstream = example_path['downstream']
log_dir = './logs/'
# load_path = "pmtransformer" (default)

# kwargs (optional)
max_epochs = 10
batch_size = 8
mean = 0
std = 1


moftransformer.run(root_dataset, downstream, log_dir=log_dir,                   
                   max_epochs=max_epochs, batch_size=batch_size,
                   mean=mean, std=std)

Test fine-tuned model

from pathlib import Path
import moftransformer
from moftransformer.examples import example_path

root_dataset = example_path['root_dataset']
downstream = example_path['downstream']

# Get ckpt file
seed = 0               # default seeds
version = 0            # version for model. It increases with the number of trains

# For version > 2.1.1, best.ckpt exists
checkpoint = 'best'    # Epochs where the model is stored. 
save_dir = 'result/'

# optional keyword
mean = 0
std = 1

load_path = Path(log_dir) / f'pretrained_mof_seed{seed}_from_pmtransformer/version_{version}/checkpoints/{checkpoint}.ckpt'

if not load_path.exists():
    raise ValueError(f'load_path does not exists. check path for .ckpt file : {load_path}')

moftransformer.test(root_dataset, load_path, downstream=downstream,
                   save_dir=save_dir, mean=mean, std=std)

predict from fine-tuned model

from pathlib import Path
import moftransformer
from moftransformer.examples import example_path

root_dataset = example_path['root_dataset']
downstream = example_path['downstream']

# Get ckpt file
log_dir = './logs/'    # same directory make from training
seed = 0               # default seeds
version = 0            # version for model. It increases with the number of trains
checkpoint = 'best'    # Epochs where the model is stored. 
mean = 0
std = 1

load_path = Path(log_dir) / f'pretrained_mof_seed{seed}_from_pmtransformer/version_{version}/checkpoints/{checkpoint}.ckpt'

if not load_path.exists():
    raise ValueError(f'load_path does not exists. check path for .ckpt file : {load_path}')
    
moftransformer.predict(
    root_dataset, load_path=load_path, downstream=downstream, split='all', mean=mean, std=std
)

Visualize analysis of feature importance for the fine-tuned model. (You should download or train fine-tuned model before visualization)

from moftransformer.visualize import PatchVisualizer
from moftransformer.examples import visualize_example_path

model_path = "examples/finetuned_bandgap.ckpt" # or 'examples/finetuned_h2_uptake.ckpt'
data_path = visualize_example_path
cifname = 'MIBQAR01_FSR'

vis = PatchVisualizer.from_cifname(cifname, model_path, data_path)
vis.draw_graph()

Architecture

It is a multi-modal pre-training Transformer encoder which is designed to capture both local and global features of porous materials.

The pre-traning tasks are as follows: (1) Topology Prediction (2) Void Fraction Prediction (3) Building Block Classification

It takes two different representations as input

Atom-based Graph Embedding : CGCNN w/o pooling layer -> local features
Energy-grid Embedding : 1D flatten patches of 3D energy grid -> global features

<img src="https://raw.githubusercontent.com/hspark1212/MOFTransformer/master/docs/source/assets/fig2.jpg" width="700")

Feature Importance Anaylsis

you can easily visualize feature importance analysis of atom-based graph embeddings and energy-grid embeddings.

%matplotlib widget
from visualize import PatchVisualizer

model_path = "examples/finetuned_bandgap.ckpt" # or 'examples/finetuned_h2_uptake.ckpt'
data_path = 'examples/visualize/dataset/'
cifname = 'MIBQAR01_FSR'

vis = PatchVisualizer.from_cifname(cifname, model_path, data_path)
vis.draw_graph()

vis = PatchVisualizer.from_cifname(cifname, model_path, data_path)
vis.draw_grid()

Universal Transfer Learning

Comparison of mean absolute error (MAE) values for various baseline models, scratch, MOFTransformer, and PMTransformer on different properties of MOFs, COFs, PPNs, and zeolites. The bold values indicate the lowest MAE value for each property. The details of information can be found in PMTransformer paper

| Material | Property | Number of Dataset | Energy histogram | Descriptor-based ML | CGCNN | Scratch | MOFTransformer | PMTransformer | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | MOF | H2 Uptake (100 bar) | 20,000 | 9.183 | 9.456 | 32.864 | 7.018 | 6.377 | 5.963 | | MOF | H2 diffusivity (dilute) | 20,000 | 0.644 | 0.398 | 0.6600 | 0.391 | 0.367 | 0.366 | | MOF | Band-gap | 20.373 | 0.913 | 0.590 | 0.290 | 0.271 | 0.224 | 0.216 | | MOF | N2 uptake (1 bar) | 5,286 | 0.178 | 0.115 | 0.108 | 0.102 | 0.071 | 0.069 | | MOF | O2 uptake (1 bar) | 5,286 | 0.162 | 0.076 | 0.083 | 0.071 | 0.051 | 0.053 | | MOF | N2 diffusivity (1 bar) | 5,286 | 7.82e-5 | 5.22e-5 | 7.19e-5 | 5.82e-05 | 4.52e-05 | 4.53e-05 | | MOF | O2 diffusivity (1 bar) | 5,286 | 7.14e-5 | 4.59e-5 | 6.56e-5 | 5.00e-05 | 4.04e-05 | 3.99e-05 | | MOF | CO2 Henry coefficient | 8,183 | 0.737 | 0.468 | 0.426 | 0.362 | 0.295 | 0.288 | | MOF | Thermal stability | 3,098 | 68.74 | 49.27 | 52.38 | 52.557 | 45.875 | 45.766 | | COF | CH4 uptake (65bar) | 39,304 | 5.588 | 4.630 | 15.31 | 2.883 | 2.268 | 2.126 | | COF | CH4 uptake (5.8bar) | 39,30

MOFTransformer

Install / Use

README