# matbench-speed

**Repository Path**: pfsuo/matbench-speed

## Basic Information

- **Project Name**: matbench-speed
- **Description**: matbench-speed
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-03-04
- **Last Updated**: 2026-03-04

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Matbench-Speed

**[View Interactive Benchmarks](https://mitkotak.github.io/matbench-speed/)**

Benchmarking MLIPs using their ASECalculator (including neighborlist, per step CPU sync overhead). All benchmark settings can be found under `scripts`

- SevenNet, GRACE, Nequix scripts out of the box.
- For MACE without kernels you might need to do [this hack](https://github.com/ACEsuit/mace/issues/1208#issuecomment-3348723384) in your local installation (typically in your conda/venv path)
- For eSEN the checkpoint needs to be downloaded from [hugging face](https://huggingface.co/facebook/OMAT24).
- For NequIP we are using the OMat24 model instead of MPTrj since it contains the kernels, and we confirmed that the hyperparams are the same for L model. XL model changes the lmax for intermediate irreps from lmax 3 to 4.