CrabNet

Overview

CrabNet, short for Compositionally-Restricted Attention-Based Network, is a software implementation of a materials-property prediction model. Its defining input is composition information: the project describes predictions made from composition alone rather than requiring additional material descriptors. The output is a predicted material property, although the supplied overview does not specify individual prediction targets, input file formats, or output schemas.

In a materials research workflow, CrabNet can be considered for the composition-to-property modeling stage. The repository accompanies two publications: a 2021 paper on the prediction model and a 2022 paper on explainable deep learning in materials science. Alongside the predictive implementation, the project demonstrates model interpretability techniques. These are presented as a separate documentation pathway, allowing readers to distinguish basic model use from interpretation-oriented work.

The main README directs users to README_CrabNet.md for installation and basic use, and to README_ExplainableGap.md for interpretability steps. Those detailed documents are referenced but were not included in the source excerpts, so dependencies, execution procedures, training requirements, supported properties, and the specific interpretation methods remain unconfirmed here. The excerpts also provide no quantitative performance results or compatibility claims. Evaluation should therefore begin with the linked instructions and a clearly defined property-prediction task, rather than assuming suitability for a particular materials dataset.

Key Features

• Implements the Compositionally-Restricted Attention-Based Network for material-property prediction.
• Uses composition information alone as the stated model input.
• Demonstrates model interpretability techniques, with a separate guide referenced for the interpretation workflow.
• Provides repository code accompanying publications on CrabNet prediction and explainable deep learning in materials science.

Use Cases

• Intended evaluation: assess composition-only property prediction for a materials dataset whose target and input representation match the detailed CrabNet instructions.
• Intended evaluation: explore the documented interpretability workflow to examine model behavior alongside property predictions.
• Use the repository as a code reference when studying the two cited CrabNet publications, while checking the detailed guides before attempting their workflows.

How to Use

1. Start with the official README to understand the composition-only prediction scope and locate the basic-use and interpretability documentation paths.
2. Follow its README_CrabNet.md link for installation and basic use. Check that guide for dependencies, input representation, and execution instructions; the supplied overview does not establish those details.
3. Define an intended evaluation dataset and target property. Confirm against the detailed instructions that the composition records and prediction task are appropriate before attempting a run.
4. Follow the README_ExplainableGap.md link from the main README if interpretation is part of your evaluation. Consult it to identify the actual methods and prerequisites rather than assuming a particular explanation technique.
5. Read the prediction paper and explainability paper for research context and citation information. Document your own evaluation conditions separately from the source-described capabilities.