An automatic "black-box" yet interpretable prediction engine for materials properties.

Automatminer is a tool for automatically creating machine learning pipelines. Automatminer's pipelines include automatic featurization with matminer, feature reduction, and AutoML backend handling. Put in a materials dataset, get out a machine that predicts materials properties.

Automatminer can make pipelines to accurately predict the properties from many kinds of materials data:

• both computational and experimental data
• small (~100 samples) to moderate (~100,000 samples) sized datasets
• crystalline datasets
• composition-only (i.e., unknown phases) datasets
• automatminer is agnostic to the target property, meaning it can be used to predict electronic, mechanical, thermodynamic, or any other kind of property

Automatminer automatically decorates a dataset using hundreds of descriptor techniques from matminer's descriptor library, picks the most useful features for learning, and runs a separate AutoML pipeline using TPOT. Once a pipeline has been fit, it can be examined with skater's interpretability tools, summarized in a text file, saved to disk, or used to make new predictions.

The easiest (and most automatic) way to use automatminer is through the MatPipe object. First, fit the MatPipe to a dataframe containing materials objects such as chemical compositions (or pymatgen Structures) and some material target property.

from automatminer.pipeline import MatPipe

# Fit a pipeline to training data to predict band gap
pipe = MatPipe()
pipe.fit(train_df, "band gap")

Now use your pipeline to predict the properties of some other data, such as a new composition or structure.

predicted_df = pipe.predict(other_df, "band gap")

You can also use it to benchmark against other machine learning models with the benchmark method of MatPipe, which optimizes the pipeline a training data and returns predictions on a held test set.

pipe = MatPipe()
test_predictions = pipe.benchmark(df, "bulk modulus", test_spec=0.2)

Once a MatPipe has been fit, you can examine it internally to see how it works using pipe.digest(); or pickle it for later with pipe.save().

We are in the process of writing a paper for automatminer. In the meantime, please use the citation given in the matminer repo.

Interested in contributing? See our contribution guidelines and make a pull request! Please submit questions, issues / bug reports, and all other communication through the matminer Google Group.