HOME > CrabNet

CrabNet

CrabNet (Compositionally-Restricted Attention-Based Network) is an open-source transformer tailored to composition-driven materials property prediction. It adapts natural-language attention blocks to encode formulas as element-distribution matrices, letting researchers train property regressors with minimal feature engineering.

Information
Case / Article
Learning Content
Review
Developer’s Voice

Information

Official site	https://crabnet.readthedocs.io
Openness	★★★
License	MIT License
Core Developers	Developers: Original authors: Anthony Wang, Steven K. Kauwe, Ryan J. Murdock, and Taylor D. Sparks. Current refactored release maintained by Sterling Baird and collaborators in the Materials Data & Informatics group at the University of Utah.
Related Papers	Anthony Yu-Tung Wang, Steven K. Kauwe, Ryan J. Murdock, Taylor D. Sparks, “Compositionally restricted attention-based network for materials property predictions,” npj Computational Materials 7, 77 (2021). doi:10.1038/s41524-021-00545-1 Sterling G. Baird et al., “Extendable composition-driven materials modeling with CrabNet,” MRS Communications 12, 1081–1089 (2022). doi:10.1557/s43579-022-00211-6
Other	Description: CrabNet packages a scikit-learn-like `fit/predict` API (`CrabNet` class) backed by PyTorch. Users feed pandas DataFrames with at least `formula` and `target` columns, and optionally pass auxiliary process variables through the `extendfeatures` mechanism. The documentation ships with lightweight datasets (`crabnet.data.materialsdata.elasticity`) plus helper utilities (`get_data`) to bootstrap experiments or teaching demos. Built-in scripts range from bare-bones educational transformers to Colab-ready notebooks that add uncertainty estimates, learning-rate cycling, and figure export (`crabnet.utils.figures`). CrabNet can be paired with Matminer/DScribe to provide feature-rich baselines that benchmark deep attention models against descriptor-based workflows for MatDaCs articles.