Skip to content

Matbench

Hacking Materials Research Group

Matbench provides 13 curated materials-science machine learning tasks for benchmarking property-prediction methods, with benchmark data, leaderboards and a pip-installable package.

Catalog updated ·

Website ↗GitHub / Repository ↗Documentation ↗
Follow

Public link：https://chemaiatlas.com/datasets/matbench/

Claim this resource

Overview

Matbench is a materials-science benchmark resource comprising 13 curated machine learning tasks. Its stated purpose is benchmarking and performance testing, and the cited paper places it in the context of materials property prediction. It is best understood as a task suite for assessing prediction methods, rather than as a predictive model that supplies its own property estimates or a materials-discovery application.

The project connects a pip-installable package with a website offering a leaderboard, per-task leaderboards, full benchmark data and a code reference for MatbenchBenchmark. These resources provide starting points for selecting a task, examining benchmark information and investigating the package interface. In a proposed evaluation workflow, a researcher would use the benchmark tasks and associated data to assess a prediction method, then consult the leaderboard resources for comparison context.

The project README does not describe individual task inputs, target properties, dataset sizes, output schemas, metrics or evaluation splits. Those details must therefore be checked in the linked documentation and benchmark pages before designing an experiment. The README states support for Python 3.8+, but the source excerpts do not establish independently tested compatibility or model performance. The repository code has an MIT licence; that evidence does not separately establish terms for every dataset or external service linked by the project.

Key Features

  • A suite of 13 curated materials-science machine learning tasks for benchmarking and performance testing.
  • A website with a general leaderboard and links to per-task leaderboards.
  • A linked full benchmark data resource for inspecting benchmark information.
  • A pip-installable matbench package, with Python 3.8+ support stated in the README.
  • Installation documentation and a code reference for MatbenchBenchmark.
  • A linked materials-science discussion forum for help and support.

Use Cases

  • Suggested evaluation: assess a materials property-prediction method on selected Matbench tasks after verifying their inputs, targets and evaluation rules.
  • Suggested evaluation: investigate differences between prediction methods using the project’s general and per-task leaderboard resources.
  • Suggested evaluation: use the linked benchmark data and code reference to plan a repeatable comparison workflow for a materials machine learning study.

How to Use

  1. Start at the repository to confirm the benchmark’s scope and locate its official resources. Treat it as an evaluation suite, not a ready-made prediction model.
  2. Consult the installation documentation. The README supplies pip install matbench and states Python 3.8+ support; check the documentation for further setup details.
  3. Examine the full benchmark data resource. Before selecting a task, verify its input representation, target, data terms and evaluation requirements.
  4. Read the MatbenchBenchmark code reference to determine how the package fits your evaluation workflow. Verify documented inputs and outputs rather than assuming an interface from the README.
  5. Consult the leaderboard and linked per-task results for comparison context. For a publication using Matbench, review the cited paper and its requested citation.

Related resources

ALIGNN

Model

ALIGNN provides atomistic graph neural networks for materials property prediction, with training workflows, pretrained predictors and ALIGNN-FF force fields for structural optimization.

Open sourcePython

Materials Discovery

Updated

Allegro

Model

Allegro implements an E(3)-equivariant interatomic potential as a NequIP extension, with documented GPU acceleration options and a separate plugin for LAMMPS simulations.

Open sourcePython

Computational Chemistry · Materials Discovery

Updated

ASE Skill (Jinzhe Zeng Group)

Skill

An ASE routing Skill in the computational-chemistry-agent-skills collection that separates workflow preparation from calculator configuration and delegates execution elsewhere.

Computational Chemistry · Materials Discovery

Updated

CGCNN

Model

CGCNN implements crystal graph convolutional neural networks for learning material properties from crystal structures, with custom-data training and prediction using pre-trained models.

Open sourcePython

Materials Discovery

Updated

ChatMOF

Agent

ChatMOF is a code-available research system that uses language-model-guided tools to retrieve MOF data, predict properties and generate structures from natural-language requests.

Open sourcePython

Materials Discovery

Updated

ChemAgent (AI4Chem)

Agent

ChemAgent (AI4Chem) is a research framework for chemistry and materials tool use, linked to the CheMatAgent paper on tree-search planning, tool execution and ChemToolBench-based training.

Computational Chemistry · Materials Discovery

Updated

Related guides

Workflows

Reproducibility for Chemistry AI Workflows

Build a compact evaluation record that connects scientific questions to data provenance, transformations, model configurations, outputs and failures, with practical planning examples for chemistry AI.

· 6 min read