ML on HPC,
without the
headache.

ml-toolkit wraps the complexity of running AI and machine learning models on HPC clusters into a single, simple command-line interface — powered by containers.

Although, Primarily designed and optimised for running AI/ML models on the Bede Grace-Hopper HPC cluster it can be easily installed on any Linux system. From a Laptop to a large HPC cluster.

See the code on GitHub → Read the Docs →
# install
$ pip install bede-ml-toolkit

# load a top-ranked ML potential
$ ml-toolkit load eSEN-30M-OAM
✓ Container image ready

# run your simulation script
$ ml-toolkit run eSEN-30M-OAM python3 simulation.py
Running inside container...
✓ Done

Features: Why ml-toolkit

01

Zero dependency wrangling

All software dependencies are packaged inside containers by the developers. You focus on your science, not your environment.

02

Portable across machines

Run the same container on your laptop or at HPC scale. Workflows move with you — no reconfiguration required.

03

GPU-ready out of the box

Containers are built to leverage GPU nodes with hardware acceleration, so you get the full power of the hardware from day one.

04

MatBench Discovery Models

Definition and config files for the all 36 models1 on the MatBench Discovery interatomic potential leaderboard are included and ready to use by name.

05

Easily extensible (not just Atomic potentials)

Can be Easily extended to add in new tools and models to work with any ML software via a simple .yaml based config.

06

Reproducible & version-locked

Software versions are frozen inside container images — your workflow won't break when external libraries update.

07

Simple, consistent CLI

Four verbs cover everything: build, run, start, stop. Whether you're running a one-off job or a long-lived background process, the interface stays the same.

08

Native Integration for ML Atomic potentials with ASE and CASTEP

Works out of the box for calculations with both the atomic simulation environment (ASE) and CASTEP.

09

Fully Open Source

Code is fully Open Source under the GPL-3.0 licence and is freely available on github.

  1. As of March 2026

Supported models

The leading ML
atomic potentials,
ready to run.

Preconfigured container definitions for the top-ranked interatomic potential models from the MatBench Discovery leaderboard. Use any model by name — no manual setup needed.

EquiformerV3+DeNS-OAM PET-OAM-XL eSEN-30M-OAM Nequip-OAM-XL MatRIS-10M-OAM SevenNet-Omni-i12 Nequip-OAM-L TACE-v1-OAM-M GRACE-2L-OAM-L ORB v3 Allegro-OAM-L + more

How it works

1

Install

One pip install command fetches the toolkit, Apptainer, and sets up your ~/ML_Toolkit directory.

2

Build a model container

ml-toolkit build <ModelName> downloads and builds the container image. Done once, reused forever.

3

Run your work

ml-toolkit run <ModelName> <cmd> executes your script inside the container with full GPU access.

N8 CIR logo

Developed with N8 Consortium funding

. ml-toolkit was developed with funding from the N8 Research Partnership through the "AI4Science" initiative (EPSRC grant number EP/T022167/1). The N8 is a consortium of the eight most research-intensive universities in the north of England, built around the N8 HPC resource, Bede. With the aim of making advanced ML capabilities accessible to researchers across all N8 institutions.