metatrain is a command line interface (CLI) to train and evaluate atomistic models of various architectures. It features a common yaml option inputs to configure training and evaluation. Trained models are exported as standalone files that can be used directly in various molecular dynamics (MD) engines (e.g. ASE, LAMMPS, i-PI, TorchSim, ESPResSo,...) using the metatomic interface.

The idea behind metatrain is to have a general training hub that provides a homogeneous environment and user interface, transforming every ML architecture into an end-to-end model that can be connected to MD engines. Any custom architecture compatible with TorchScript can be integrated into metatrain, gaining automatic access to a training and evaluation interface, as well as compatibility with various MD engines.

List of Implemented Architectures

Currently metatrain supports the following architectures for building an atomistic model:

Name	Description
PET	Point Edge Transformer (PET), interatomic machine learning potential
SOAP-BPNN	A Behler-Parrinello neural network with SOAP features
MACE	A higher order equivariant message passing neural network.
PhACE	SO(3)-equivariant message-passing model with physical radial functions and fast tensor products.
GAP	Sparse Gaussian Approximation Potential (GAP) using Smooth Overlap of Atomic Positions (SOAP).
FlashMD	An architecture for the direct prediction of molecular dynamics

Documentation

For details, tutorials, and examples, please visit our documentation.

Installation

Install metatrain with pip:

pip install metatrain

Install specific models by specifying the model name. For example, to install the SOAP-BPNN model:

pip install metatrain[soap-bpnn]

We also offer a conda installation:

conda install -c conda-forge metatrain

⚠️ The conda installation does not install model-specific dependencies and will only work for architectures without optional dependencies such as PET.

After installation, you can use mtt from the command line to train your models!

Quickstart

To train a model, use the following command:

mtt train options.yaml

Where options.yaml is a configuration file specifying training options. For example, the following configuration trains a SOAP-BPNN model on the QM9 dataset:

# architecture used to train the model
architecture:
  name: soap_bpnn
  training:
    num_epochs: 5  # a very short training run

# Mandatory section defining the parameters for system and target data of the training set
training_set:
  systems: "qm9_reduced_100.xyz"  # file where the positions are stored
  targets:
    energy:
      key: "U0"      # name of the target value
      unit: "eV"     # unit of the target value

test_set: 0.1        # 10% of the training_set are randomly split for test
validation_set: 0.1  # 10% of the training_set are randomly split for validation

Shell Completion

metatrain comes with completion definitions for its commands for bash and zsh. You must manually configure your shell to enable completion support.

To make the completions available, source the definitions in your shell’s startup file (e.g., ~/.bash_profile, ~/.zshrc, or ~/.profile):

source $(mtt --shell-completion)

Having problems or ideas?

Having a problem with metatrain? Please let us know by submitting an issue.

Submit new features or bug fixes through a pull request.

Contributors

Thanks goes to all people who make metatrain possible:

The overall metatrain project is maintained by @frostedoyster, @pfebrer, and @PicoCentauri who will reply to issues and pull requests opened on this repository as soon as possible. You can mention them directly if you did not receive an answer after a couple of days. Additionally, different architectures are maintained by separate maintainers, you can find their names in the corresponding documentation

Citing metatrain

If you found metatrain useful for your work, please cite the corresponding article:

F. Bigi, J.W. Abbott, P. Loche et. al.
Metatensor and metatomic: foundational libraries for interoperable atomistic machine learning, (2026).
https://doi.org/10.1063/5.0304911

@article{bigi_metatensor_2026,
  title = {Metatensor and Metatomic: {{Foundational}} Libraries for Interoperable Atomistic Machine Learning},
  shorttitle = {Metatensor and Metatomic},
  author = {Bigi, Filippo and Abbott, Joseph W. and Loche, Philip and Mazitov, Arslan and Tisi, Davide and Langer, Marcel F. and Goscinski, Alexander and Pegolo, Paolo and Chong, Sanggyu and Goswami, Rohit and Febrer, Pol and Chorna, Sofiia and Kellner, Matthias and Ceriotti, Michele and Fraux, Guillaume},
  year = 2026,
  month = feb,
  journal = {J. Chem. Phys.},
  volume = {164},
  number = {6},
  pages = {064113},
  issn = {0021-9606},
  doi = {10.1063/5.0304911},
}