Table of Contents

• Introduction
• SCANN Framework
• Installation
• Usage
• Datasets
• References

Introduction

This repository is the official implementation of Towards understanding structure–property relations in materials with interpretable deep learning.

Please cite us as

@article{Vu2023,
doi = {10.1038/s41524-023-01163-9},
issn = {2057-3960},
journal = {npj Computational Materials},
number = {1},
pages = {215},
title = {{Towards understanding structure–property relations in materials with interpretable deep learning}},
url = {https://doi.org/10.1038/s41524-023-01163-9},
volume = {9},
year = {2023}
}

We developed a Self-Consistent Atention-based Neural Network (SCANN) that takes advantage of a neural network to quantitatively capture the contribution of the local structure of material properties.

The model captures information on atomic sites and their local environments by considering self-consistent long-range interactions to enrich the structural representations of the materials. A comparative experiment was performed on benchmark datasets QM9 and Material Project (2018.6.1) to compare the performance of the proposed model with state-of-the-art representations in terms of prediction accuracy for several target properties.

Furthermore, the quantitative contribution of each local structure to the properties of the materials can help understand the structural-property relationships of the materials.

SCANN framework

The Self-Consistent Atention-based Neural Network (SCANN) is an implementation of deep attention mechanism for materials science.

Figure 1 shows the overall schematic of the model

Figure 1. Schematic of SCANN.

Installation

Firstly, create a conda environment to install the package, for example:

conda create -n test python==3.9
source activate test

Optional GPU dependencies

For hardwares that have CUDA support, the tensorflow version with gpu options should be installed. Please follow the installation from https://www.tensorflow.org/install for more details.

Tensorflow can also be installed from conda for simplification settings:

conda install -c conda-forge tensorflow-gpu

Method 1 (directly install from git)

You can install the lastes development version of SCANN from this repo and install using:

git clone https://github.com/sinhvt3421/scann-material
cd scann-material
python -m pip install -e .

Method 2 (using pypi)

SCANN can be installed via pip for the latest stable version:

pip install scann-model

Usage

Our current implementation supports a variety of use cases for users with different requirements and experience with deep learning. Please also visit the notebooks directory for Jupyter notebooks with more detailed code examples.

Below is an example of predicting the "HOMO" and corresponding global attention score:

from scann.utils import load_file, prepare_input_pmt
from scann.models import SCANN
import yaml

#load config and pretrained model from folders

config = yaml.safe_load(open('trained_model/homo/config.yaml'))
scann = SCANN(config, pretrained='trained_model/homo/model_homo.h5', mode='infer')

#load file for structure using pymatgen Structure 

struct = load_file('abc.xyz') # pymatgen.core.Structure 
inputs = prepare_input_pmt(struct, d_t=4.0, w_t=0.4, angle=False)  # Distance, weights threshold

# Predict the target property with the ga score for interpretation
pre_target, ga_score = scann.model.predict(inputs)

Using pre-built models

In our work, we have already built models for the QM9 [1] and Material Project 2018 [2] datasets . The model is provided as serialized HDF5+yaml files.

• QM9 molecule data:
  • HOMO: Highest occupied molecular orbital energy
  • LUMO: Lowest unoccupied molecular orbital energy
  • Gap: energy gap
  • α: isotropic polarizability
  • Cv: heat capacity at 298 K

The MAEs on the various models are given below:

Performance on QM9

Property	Units	SCANN	SCANN+
HOMO	meV	41	32
LUMO	meV	37	31
Gap	meV	61	52
α	Bohr^3	0.141	0.115
Cv	cal/(molK)	0.050	0.041

Performance on Material Project 2018.6.1

Property	Units	SCANN	SCANN+
Ef	meV(atom)-1	29	28
Eg	meV	260	225

Datasets

Experiments

The settings for experiments specific is placed in the folder configs

We provide an implementation for the QM9 experiments, the fullerene-MD, the Pt/graphene-MD, Material Project 2018.6.1, and SmFe12-CD [3] experiments.

Basic usage

Data preprocessing

For training new model for each datasets, please follow the below example scripts. If the data is not avaiable, please run the code preprocess_data.py for downloading and creating suitable data formats for SCANN model. For example:

$ python preprocess_data.py qm9 processed_data --dt=4.0 --wt=0.4 --p=8

-----

$ python preprocess_data.py mp2018 processed_data --dt=6.0 --wt=0.4 --p=8

The data for QM9 or Material Project 2018 will be automatically downloaded and processed into folder propessed_data. For all avaiable datasets and options for cutoff distance/Voronoi angle, please run python preprocess.py --h to show all details.

Model training

After that, please change the config file located in folder configs for customizing the model hyperparameters or data loading/saving path.

$ python train.py homo configs/model_qm9.yaml --use_drop=True

For training dataset fullerene-MD with pretrained weights from QM9 dataset, please follow these steps. The pretrained model will be load based on the path from argument.

$ python train.py homo configs/model_fullerene.yaml --pretrained=.../qm9/homo/models/model.h5

For running the evaluation from pretrained weights, please follow these steps.

$ python train.py homo ..../qm9/homo/configs.yaml --pretrained=.../qm9/homo/models/model.h5  --mode=eval 

Model inference

The code predict_files.py supports loading a xyz file and predicting the properties with the pretrained models. The information about global attention (GA) score for interpreting the structure-property relationship is also provided and saved into xyz format. Please use a visualization tool such as Ovito [4] for showing the results.

$ python predict_files.py ..../models.h5 save_path.../ experiments/molecules/Dimethyl_fumarate.xyz

Figure 2. Example of SCANN prediction for LUMO property.

References

[1] Ramakrishnan, R., Dral, P., Rupp, M. et al. Quantum chemistry structures and properties of 134 kilo molecules. Sci Data 1, 140022 (2014). https://doi.org/10.1038/sdata.2014.22

[2] Xie, T. & Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Phys. Rev. Lett. 120, 145301 (2018). https://doi.org/10.1021/acs.chemmater.9b01294

[3] Nguyen, DN., Kino, H., Miyake, T. et al. Explainable active learning in investigating structure–stability of SmFe_12-α-β X_αY_β structures X, Y {Mo, Zn, Co, Cu, Ti, Al, Ga}. MRS Bulletin 48, 31–44 (2023). https://doi.org/10.1557/s43577-022-00372-9

[4] A. Stukowski, Visualization and Analysis of Atomistic Simulation Data with OVITO–the Open Visualization Tool, Model. Simul. Mater. Sci. Eng. 18, 15012 (2009). https://doi.org/10.1088/0965-0393/18/1/015012