MODNet: Material Optimal Descriptor Network

Table of contents

• Introduction
• How to install
• Usage
• Pretrained models
• Stored MODData
• Documentation
• Getting started
  • MODData
  • MODNetModel
• Author
• License

Introduction

This repository contains the Python (3.8) package implementing the Material Optimal Descriptor Network (MODNet). It is a supervised machine learning framework for learning material properties from either the composition or crystal structure. The framework is well suited for limited datasets and can be used for learning multiple properties together by using joint learning.

This repository also contains two pretrained models that can be used for predicting the refractive index and vibrational thermodynamics from any crystal structure.

See the MODNet papers and repositories below for more details:

• Machine learning materials properties for small datasets, De Breuck et al. (2020), arXiv:2004.14766.

• Robust model benchmarking and bias-imbalance in data-driven materials science: a case study on MODNet, De Breuck et al. (2021), arXiv:2102.02263.

• Matbench benchmarking data repository: ml-evs/modnet-matbench.

How to install

MODNet can be installed via pip:

pip install modnet

Usage

The MODNet package is built around two classes: MODData and MODNetModel.

The usual workflow is as follows:

from modnet.preprocessing import MODData
from modnet.models import MODNetModel

# Creating MODData
data = MODData(materials = structures,
               targets = targets,
              )
data.featurize()
data.feature_selection(n=200)

# Creating MODNetModel
model = MODNetModel(target_hierarchy,
                    weights,
                    num_neurons=[[256],[64,64],[32]],
                    )
model.fit(data)

# Predicting on unlabeled data
data_to_predict = MODData(new_structures)
data_to_predict.featurize()
df_predictions = model.predict(data_to_predict) # returns dataframe containing the prediction on new_structures

Example notebooks and short tutorials can be found in the example_notebooks directory.

Pretrained models

Two pretrained models are provided in pretrained/:

• Refractive index
• Vibrational thermodynamics

Download these models locally to path/to/pretrained/. Pretrained models can then be used as follows:

from modnet.models import MODNetModel

model = MODNetModel.load('path/to/pretrained/refractive_index')
# or MODNetModel.load('path/to/pretrained/vib_thermo')

Stored MODData

The following MODDatas are available for download:

• Formation energy on Materials Project (June 2018), on figshare
• Refractive index (upon request)
• Vibrational thermodynamics (upon request)

Download this directory localy to path/to/moddata/. These can then be used as follows:

from modnet.preprocessing import MODData

data_MP = MODData.load('path/to/moddata/MP_2018.6')

The MP MODData on figshare (MP_2018.6) is very usefull for predicting a learned property on all structures from the Materials Project:

predictions_on_MP = model.predict(data_MP)

Documentation

The documentation is available at https://modnet.readthedocs.io.

Especially, carefully read the two main classes, MODData and MODNetModel found in preprocessing and models modules. A MODData instance is used for representing a particular dataset. It contains a list of structures and corresponding properties. A MODNetModelinstance is used for training and predicting of one or more properties or classes.

Getting started

MODData

A MODData instance is used for representing a particular dataset. It contains a list of structures and corresponding properties:

from modnet.preprocessing import MODData

data = MODData(
    materials,
    targets = None,
    target_names = None,
    mpids = None,
)

main arguments:

• materials (List): List of pymatgen Compositions or Structures.
• targets (List)(optional): List of targets corresponding to each structure. When learning on multiple targets this is a ndarray where each column corresponds to a target, i.e. of shape (n_materials,n_targets).
• target_names (Iterable) (optional): Iterable (e.g list) of names corresponding to the properties. E.g. ['S_300K','S_800K',...] or ['refractive_index'] for single target learning. These names are used when building the model.
• structure_ids (Iterable) (optional): If the list of structures (materials) are from the Materials Project, you can specify the corresponding mpids by providing an Iterable of mpids: ['mp-149','mp-166',...]. This will enable fast featurization (see further).

The next step is to create the features:

data.featurize(fast = False, db_file = 'feature_database.pkl')

main arguments:

• fast (Boolean) (optional): If set to True, the algorithm will use the pre-computed features from a database instead of computing them again from scratch. This is recommended (and only possible) when using structures from the Materials Project. Note that the mpids should be provided in the MODData.
• db_file (Boolean) (optional): When setting fast to True, you also need to provide this argument. Download the file at this figshare link, unzip it, then set the local path to this file as argument.

See Documentation for the complete set of arguments.

Finally, the optimal features are computed:

data.feature_selection(n=300)

main arguments:

• n (optional): Number of optimal features to compute, i.e. the n first ranked features are computed. When set to -1, all features are ranked (recommended, but can take time).

See Documentation for the complete set of arguments.

The MODData can be saved,

data.save('path/dataname')

and loaded for later usage:

from modnet.preprocessing import MODData

data = MODData.load('path/dataname')

Both the save and load methods use pandas .read_pickle(...) and .load_pickle(...) which will try to compress/decompress files according to their file extensions (e.g. ".zip", ".tgz" and ".bz2"). Please note that this method is unsafe as it can load arbitrary Python objects. Care has been taken to check the hashes of data that is automatically loaded from Figshare; if you rely on this feature for your own data then we recommend you do the same.

Dataframes for features, targets and other data can be accessed trough the following methods:

# dataframe containing the structures
data.get_structure_df()

# dataframe containing the targets
data.get_target_df()

# dataframe containing the features
data.get_featurized_df()

# List of the optimal features, in ranked order
data.get_optimal_descriptors()
    
# get_featurized_df limited to the best features
data.get_optimal_df()

MODNetModel

The model is created by a MODNetModel instance:

from modnet.models import MODNetModel

model = MODNetModel(
    targets,
    weights,
    num_classes = None  # only needed for classification, e.g. num_classes = {'is_metal':2}
    num_neurons=[[64],[32],[16],[16]],
    n_feat=300,
    act='relu',
)

main arguments:

• targets (List): Specifies how the different targets are organized in the architecture. It is a list of lists of lists, representing the three modular last levels: block 2, 3 and 4 (see Figure 2). Each block gathers properties, which are put inside the same list. For exmaple, in Figure 2, this is [[['S_5,...,S_800'],['U_5,...,U_800'],['C_v_5,...,C_v_800'],['H_5,...,H_800']],[['formation_energy']]]. The same names as given in MODData should be used.

• weights (Dictionary): A dictionary where each key is a property name and the value the corresponding weight to be used in the loss function. The weights are used to scale the different outputs such that the balance between the properties is conserved when training. For example, {'S_5':0.01, 'formation_energy:1'}.

• num_classes`: Dictionary defining the target types (classification or regression). Should be constructed as follows: key: string giving the target name; value: integer n, with n=0 for regression and n>=2 for classification with n the number of classes.

• num_neurons (List) (optional): Number of neurons as well as the number of layers to be used in the neural network. List of three lists. Each inner list gives respectively the succesive number of neurons of the blocks 2, 3 and 4. For example, in Figure 2, this is given by [[128,128],[64,64],[8]].

• n_feat (int) (optional): Number of optimal features to be used in the model. In Figure 2, this is 330.

• act (String) (optional): Activation function used in the neural network, see Keras API.

See Documentation for the complete set of arguments.

The model is then fitted on the data:

model.fit(
    data,
    val_fraction = 0.0,
    val_data = None,
    val_key = None,
    lr=0.001,
    epochs = 200,
    batch_size = 128,
    xscale='minmax',
    loss='mse',
)

main arguments:

• val_fraction (float) (optional): Validation fraction to be used while training.
• val_data (float) (optional): Validation MODData to be used while training.
• val_key (String) (optional): The name of the property used for printing validation MAE. When multiple properties are learned (e.g. ['formation energy', 'refractive_index', 'entropy']), setting the key_val (e.g. key_val = 'entropy') will only print the MAE of this property for each epoch.
• lr (float) (optional): Learning rate.
• epochs (int) (optional): Number of epochs.
• batch_size (int) (optional): Batch size.
• xscale (String) (optional): Scaling of the features. Possible values: 'minmax' or 'standard'.
• loss (String)(optional): Loss function of the neural network, see Keras API.

See Documentation for the complete set of arguments.

You can save and load the model for later usage:

model.save('path/modelname')

from modnet.models import MODNetModel
MODNetModel.load('path/modelname')

Prediction is done by first creating a MODData instance on the new data:

data_to_predict = MODData(materials, mpids = df.index) # Adding mpids is a good idea for fast featurization if possible, but not necessary.
data_to_predict.featurize(fast=True)

and then using the predict method:

df_predictions = model.predict(data_to_predict)

A dataframe containing the predictions is returned.

Author

The first versions of this software were written by Pierre-Paul De Breuck, with contributions from Matthew Evans (v0.1.7+).

License

MODNet is released under the MIT License.