maml (MAterials Machine Learning) provides high-level interfaces that make machine learning for materials science as easy as possible. It does not reimplement ML algorithms. Instead, it builds on scikit-learn, TensorFlow and MatGL for models, and on pymatgen and matminer for structure handling and featurization.

• Documentation: https://materialyzeai.github.io/maml/
• API reference: https://materialyzeai.github.io/maml/maml.html
• Example notebooks: notebooks
• Tutorial lecture and hosted tool: nanoHUB

Features

Module	What it does
maml.describers	Convert compositions, sites and structures into features: element statistics, Coulomb matrices, radial distribution functions, bispectrum coefficients, Behler-Parrinello symmetry functions, SOAP, and graph-network (MatGL, MEGNet, M3GNet) embeddings. Any matminer featurizer can be wrapped with wrap_matminer_describer.
maml.models	Describer-aware wrappers around scikit-learn (SKLModel) and Keras (KerasModel, MLP) models, plus the AtomSets transfer-learning architecture.
maml.sampling	DIRECT (DImensionality REduction-Clustering-sTratified) sampling for selecting robust, diverse training sets from large configuration spaces.
maml.apps.pes	Fit and evaluate machine-learning interatomic potentials: SNAP/qSNAP, GAP, MTP and NNP, with LAMMPS-based property calculators (elastic constants, lattice constants, defect formation, NEB).
maml.apps.bowsr	BOWSR: Bayesian optimization with a surrogate energy model for DFT-free structure relaxation.
maml.apps.gbe	Machine-learning model for elemental grain boundary energies.
maml.apps.symbolic	Symbolic regression and sparse feature selection: SIS/ISIS, LASSO, adaptive LASSO, SCAD, Dantzig selector and L0 search.

Installation

pip install maml

Or, with uv:

uv add maml

Optional extras:

pip install "maml[symbolic]"  # cvxpy, for maml.apps.symbolic
pip install "maml[deep]"      # TensorFlow, for Keras models and AtomSets

maml.apps.pes drives external codes, which must be installed separately and be on your PATH:

Potential	Required code
All potentials (property calculations)	LAMMPS (conda install -c conda-forge lammps)
SNAP / qSNAP	LAMMPS (built in)
GAP	QUIP with the GAP package, and LAMMPS built with ML-QUIP
MTP	MLIP-3 (default; MLIP-2 via MTPotential(version="mlip-2")). LAMMPS-based MTP calculations currently require MLIP-2 and its LAMMPS interface
NNP	n2p2

Development

git clone https://github.com/materialyzeai/maml.git
cd maml
uv sync --group lint
uv run pytest tests

Quick start

Train a composition model with scikit-learn:

from pymatgen.core import Composition
from sklearn.linear_model import LinearRegression

from maml.describers import ElementStats
from maml.models import SKLModel

describer = ElementStats.from_data("megnet_3", stats=["mean", "std"])
model = SKLModel(describer=describer, model=LinearRegression())

compositions = [Composition(c) for c in ["Li2O", "Na2O", "K2O", "MgO", "CaO"]]
targets = [-2.07, -1.40, -1.13, -3.08, -3.27]  # formation energies, eV/atom
model.train(compositions, targets)
print(model.predict_objs([Composition("SrO")]))

See the notebooks for worked examples of each module.

Citing

If you use maml, please cite:

@misc{maml,
    author = {Chen, Chi and Zuo, Yunxing and Ye, Weike and Qi, Ji and Ong, Shyue Ping},
    title = ,
    year = {2020},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/materialyzeai/maml}},
}

Please also cite the paper for the module you use:

• ML-IAPs (maml.apps.pes): Zuo, Y.; Chen, C.; Li, X.; Deng, Z.; Chen, Y.; Behler, J.; Csányi, G.; Shapeev, A. V.; Thompson, A. P.; Wood, M. A.; Ong, S. P. Performance and Cost Assessment of Machine Learning Interatomic Potentials. J. Phys. Chem. A 2020, 124 (4), 731–745. https://doi.org/10.1021/acs.jpca.9b08723
• BOWSR (maml.apps.bowsr): Zuo, Y.; Qin, M.; Chen, C.; Ye, W.; Li, X.; Luo, J.; Ong, S. P. Accelerating Materials Discovery with Bayesian Optimization and Graph Deep Learning. Mater. Today 2021, 51, 126–135. https://doi.org/10.1016/j.mattod.2021.08.012
• AtomSets (maml.models.AtomSets): Chen, C.; Ong, S. P. AtomSets as a Hierarchical Transfer Learning Framework for Small and Large Materials Datasets. npj Comput. Mater. 2021, 7, 173. https://doi.org/10.1038/s41524-021-00639-w
• DIRECT sampling (maml.sampling): Qi, J.; Ko, T. W.; Wood, B. C.; Pham, T. A.; Ong, S. P. Robust Training of Machine Learning Interatomic Potentials with Dimensionality Reduction and Stratified Sampling. npj Comput. Mater. 2024, 10, 43. https://doi.org/10.1038/s41524-024-01227-4
• Grain boundary energies (maml.apps.gbe): Ye, W.; Zheng, H.; Chen, C.; Ong, S. P. A Universal Machine Learning Model for Elemental Grain Boundary Energies. Scr. Mater. 2022, 218, 114803. https://doi.org/10.1016/j.scriptamat.2022.114803