Matminer

Warnings

• breaking In `matminer v0.10.0`, the `impute_nan` parameter for many featurizers changed its default value from `False` to `True`. This means missing features (NaNs) will now be automatically imputed as the mean of the column by default, potentially altering previous model training behaviors if not explicitly set to `False`.
  Fix: If you require the old behavior (no imputation by default), explicitly set `impute_nan=False` when initializing your featurizers, e.g., `MyFeaturizer(impute_nan=False)`.
• gotcha Matminer has strict dependency requirements, particularly for `pandas`, `numpy`, and `pymatgen`. Incompatibilities can arise with newer versions of these upstream libraries, leading to import errors or unexpected behavior. This was particularly noted around `pandas v2` and specific `pymatgen` versions.
  Fix: If encountering dependency issues, consider installing matminer in a fresh environment. Refer to the `requirements/*.txt` files on the GitHub repository for pinned dependency versions known to be compatible with your matminer version. Always check release notes for specific compatibility updates.
• gotcha When using `BaseFeaturizer.set_n_jobs()` for parallel processing, especially with large datasets, there is a risk of Out-of-Memory (OOM) errors. This is due to Python's multiprocessing overhead and potential data duplication across processes if not managed carefully (e.g., with a Dask client).
  Fix: For very large datasets, avoid setting `n_jobs` to a high number directly. Consider processing in chunks, using a Dask client for more robust distributed computing, or setting `n_jobs=1` to run in a single process if memory is a constraint.
• deprecated In `matminer v0.10.0`, the `ChemEnvSiteFingerprint.from_preset()` method had some 'not-implemented' Chemical Environments (CEs) removed. While these were technically non-functional, their removal might affect code that expected a specific set of presets or relied on the method's previous behavior.
  Fix: Review your usage of `ChemEnvSiteFingerprint.from_preset()`. Ensure that the specific Chemical Environments you intend to use are still available. If you were using any of the removed 'not-implemented' CEs, adapt your code to use the currently supported ones or implement custom CEs if needed.

Install

• pip install matminer Latest stable release

Imports

• load_elastic_debye

  from matminer.datasets.dataframe_loader import load_elastic_debye

• StrToStructure

  from matminer.featurizers.conversions import StrToStructure

• ElementProperty

  from matminer.featurizers.composition import ElementProperty

• SiteStatsFingerprint

  from matminer.featurizers.structure import SiteStatsFingerprint

Quickstart

This quickstart demonstrates how to load a dataset from matminer, convert string representations of structures into pymatgen Structure objects, and apply a composition-based featurizer to enrich the DataFrame with new material properties.

import pandas as pd
from matminer.datasets.dataframe_loader import load_elastic_debye
from matminer.featurizers.conversions import StrToStructure
from matminer.featurizers.composition import ElementProperty

# 1. Load a sample dataset
df = load_elastic_debye()
print(f"Initial DataFrame shape: {df.shape}")
print(df.head())

# 2. Convert 'structure' string representation to pymatgen Structure objects
# Note: This step is crucial for structure-based featurizers
str_to_structure = StrToStructure()
df = str_to_structure.featurize_dataframe(df, 'structure')

# 3. Apply a composition featurizer
# (ElementProperty calculates statistics of elemental properties for each composition)
ep_featurizer = ElementProperty(
    features=['atomic_radius', 'electronegativity'],
    stats=['mean', 'std_dev']
)
df = ep_featurizer.featurize_dataframe(df, 'composition', ignore_errors=True)

print(f"\nDataFrame after featurization shape: {df.shape}")
print(df[['formula', 'composition', 'ElementProperty mean atomic_radius', 'ElementProperty std_dev electronegativity']].head())