CatBench v1.0.0 - Cell Reports Physical Science

Publication Release

This release corresponds to the version used in our paper accepted at Cell Reports Physical Science: "CatBench Framework for Benchmarking Machine Learning Interatomic Potentials in Adsorption Energy Predictions for Heterogeneous Catalysis"

Key Features

• Benchmarking framework for MLIP adsorption energy predictions in heterogeneous catalysis
• Multi-class anomaly detection system (normal, adsorbate migration, anomaly)
• Support for Catalysis-Hub datasets and user-generated DFT data
• Comprehensive evaluation metrics (MAE, classification rates, ADwT, AMDwT)
• Compatible with 13+ state-of-the-art MLIPs including UMA, GRACE, eSEN, CHGNet, MACE, SevenNet, ORB

Benchmaring Performance

• Evaluated on 45,130 small molecule adsorption reactions (MamunHigh2019 dataset)
• Evaluated on 2,651 large molecule adsorption reactions (FG dataset)
• Best models achieve normal MAE ~0.2 eV for small molecules and ~0.13 eV for large molecules with dispersion corrections

Installation and Usage

See README for detailed installation instructions and usage examples.

Citation

If you use CatBench in your research, please cite our Cell Reports Physical Science paper

Authors

Jinuk Moon, Uchan Jeon, Seokhyun Choung, Jeong Woo Han
Department of Materials Science and Engineering, Seoul National University

---

Note: This is the stable version used for all results reported in the paper.