Matbench Discovery: Inorganic Crystal Stability Prediction Data

Matbench Discovery v1 provides data files for a benchmark on machine learning energy models predicting inorganic crystal stability from unrelaxed structures. The dataset includes relaxed structures from the Materials Project training set and initial plus relaxed structures from the WBM test set, along with model checkpoints. It was authored by Janosh Riebesell and updated in May 2026.

Use Cases

• Benchmarking ML model performance on crystal stability prediction based on the described test and training sets.
• Training energy models using the provided relaxed structures, energies, forces, and stresses.
• Analyzing model checkpoints to understand training progression for this specific benchmark task.

Strengths

• Dataset size is 15.6 GB, indicating a substantial collection of structural data.
• Includes both initial and relaxed crystal structures for the WBM test set, providing a trajectory for analysis.
• Relies on established sources like the Materials Project, which contains 1.3 million structures.

Limitations

• Row count for the primary benchmark files is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download from the linked description.

Provenance

Source

Materials Project (MP) training set and WBM test set.

Collection Method

Computational relaxation of inorganic crystal structures.

Freshness

Last updated 2026-05-19 11:57:38.