Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry
Tsz Wai Ko, Bowen Deng, Marcel Nassar, Luis Barroso-Luque, Runze Liu, Ji Qi, Elliott Liu, Gerbrand, Ceder, Santiago Miret, Shyue Ping Ong · 2025
Read Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry by Tsz Wai Ko, Bowen Deng, Marcel Nassar, Luis Barroso-Luque, Runze Liu, Ji Qi, Elliott Liu, Gerbrand, Ceder, Santiago Miret, Shyue Ping Ong (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-03837.txt.
Introduction
Tsz Wai Ko leads a team of ten authors in 2025 in releasing MatGL. MatGL is an open-source library that brings graph deep learning to materials science and chemistry. Atoms in a material link to their neighbors like nodes in a web. Graph models suit that structure naturally.
The library builds on the Deep Graph Library and the Pymatgen toolkit. It aims to be batteries-included: ready models, data handling, and training tools in one place, extensible for specialists. Such libraries matter. They let a small lab reuse machinery that once needed a dedicated team, speeding the search for better batteries, catalysts, and alloys.
Catalog record
| CREATOR | Tsz Wai Ko, Bowen Deng, Marcel Nassar, Luis Barroso-Luque, Runze Liu, Ji Qi, Elliott Liu, Gerbrand, Ceder, Santiago Miret, Shyue Ping Ong |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 70 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-03837
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry Tsz Wai Ko, Bowen Deng, Marcel Nassar, Luis Barroso-Luque, Runze Liu, Ji Qi, Elliott Liu, Gerbrand, Ceder, Santiago Miret, Shyue Ping Ong Abstract Graph deep learning models, which incorporate a natural inductive bias for a collection of atoms, are of immense interest in materials science and chemistry. Here, we introduce the Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry. Built on top of the popular Deep Graph Library (DGL) and Python Materials Genomics (Pymatgen) packages, our intention is for MatGL to be an extensible “batteries-included” library for the development of advanced graph deep learning models for materials property predictions and interatomic potentials. At present, MatGL has efficient implementations for both invariant and equivariant graph deep learning models, including the Materials 3-body Graph Network (M3GNet), MatErials Graph Network (MEGNet), Crystal Hamiltonian Graph Network (CHGNet), TensorNet and SO3Net architectures. MatGL also includes a variety of pre-trained universal interatomic potentials (aka “foundational materials models (FMM)”) and property prediction models are also included for out-of-box usage, benchmarking and fine-tuning. Finally, MatGL includes support for Pytorch Lightning for rapid training of models. † † affiliation: Aiiso Yufeng Li Family Department of Chemical and Nano Engineering, University of California San Diego, 9500 Gilman Dr, Mail Code 0448, La Jolla, CA 92093-0448, United States † † email: [email protected] † † affiliation: Department of Materials Science and Engineering, University of California Berkeley, Berkeley, CA, USA † † affiliation: Materials Sciences Division, Lawrence Berkeley National Laboratory, California 94720, United States † † affiliation: Intel Labs, Santa Clara, CA, United States † † affiliation: Department of Materials Science and Engineering, University of California Berkeley, Berkeley, CA, USA † † affiliation: Materials Sciences Division, Lawrence Berkeley National Laboratory, California 94720, United States † † affiliation: Aiiso Yufeng Li Family Department of Chemical and Nano Engineering, University of California San Diego, 9500 Gilman Dr, Mail Code 0448, La Jolla, CA 92093-0448, United States † † affiliation: Aiiso Yufeng Li Family Department of Chemical and Nano Engineering, University of California San Diego, 9500 Gilman Dr, Mail Code 0448, La Jolla, CA 92093-0448, United States † † affiliation: Aiiso Yufeng Li Family Department of Chemical and Nano Engineering, University of California San Diego, 9500 Gilman Dr, Mail Code 0448, La Jolla, CA 92093-0448, United States † † affiliation: Department of Materials Science and Engineering, University of California Berkeley, Berkeley, CA, USA † † affiliation: Materials Sciences Division, Lawrence Berkeley National Laboratory, California 94720, United States † † affiliation: Intel Labs, Santa Clara, CA, United States † † email: [email protected] † † affiliation: Aiiso Yufeng Li Family Department of Chemical and Nano Engineering, University of California San Diego, 9500 Gilman Dr, Mail Code 0448, La Jolla, CA 92093-0448, United States 1 Introduction In recent years, machine learning (ML) has emerged as a powerful new tool in the materials scientist’s toolkit Chen et al. (2020) ; Schmidt et al. (2019) ; Westermayr et al. (2021) ; Oviedo et al. (2022) . Sophisticated ML models have found their way into a multitude of applications. Surrogate ML models for “instant” predictions of properties such as formation energies, band gaps, mechanical properties, etc. Chen et al. (2019) ; Schmidt et al. (2021) ; Gasteiger et al. (2020) ; Gasteiger et al. (2020) ; Satorras et al. (2021) ; Liu et al. (2022) ; Brandstetter et al. (2022) ; Kaba and Ravanbakhsh (2022) ; Yan et al. (2022) have greatly expanded our ability to explore vast chemical spaces for new materials. In addition, Machine learning (ML) has been widely used for parameterizing potential energy surfaces (PESs), enabling the direct prediction of potential energies, forces, and stresses based on atomic positions and chemical species. These ML interatomic potentials (MLIPs) Schütt et al. (2018) ; Schütt et al. (2021) ; Bartók et al. (2010) ; Behler and Parrinello (2007) ; Thompson et al. (2015) ; Drautz (2019) ; Batzner et al. (2022) ; Ko et al. (2023) ; Kocer et al. (2022) ; Ko et al. (2021) ; Ko and Ong (2023) ; Unke et al. (2021) ; Liao and Smidt (2023) have provided us with the means to parameterize complex PESs to perform large-scale atomistic simulations with unprecedented accuracies. Fig. 1 : Graph deep learning architecture for materials science. \mathbf{V_{n}} and \mathbf{E_{n}} denotes the set of node/atom ( \{\mathbf{v_{i}}\} ) and edge/bond features ( \{\mathbf{e_{ij}}\} ), respectively, in the n^{th} layer. Some implementations include a global state feature ( \mathbf{U} ) for greater expressive power. Between layers, a sequence of edge ( f_{E} ), node ( f_{V} ) and state ( f_{U} ) update operations are performed. f_{E} , f_{V} and f_{U} are usually modeled using multilayer perceptrons. In the final step, the edges, nodes and state features are pooled ( P ) and passed through a multilayer perceptron to arrive at a prediction. Among ML model architectures, graph deep learning models, also known as graph neural networks (GNNs), utilize a natural representation that incorporates a physically intuitive inductive bias for a collection of atoms. Battaglia et al. (2018) Figure 1 depicts a typical graph deep learning architecture. In the graph representation, the atoms are nodes and the bonds between atoms (usually defined based on a cutoff radius) are edges. In most implementations, each node is represented by a learned embedding vector for each unique atom type (element). Additionally, some architectures such as the MatErials Graph Network (MEGNet) Chen et al. (2019) and Materials 3-body Graph Network (M3GNet) Chen and Ong (2022) also include an optional global state feature ( \mathbf{u} ) to provide greater expressive power, for instance, in the handling of multifidelity data. Chen et al. (2021) ; Ko and Ong (2024) A graph deep learning model is constructed by performing a sequence of update operations, also known as message passing or graph convolutions. In the final layer, the embeddings are pooled and passed through a final MLP layer to arrive at a final prediction. GNNs can be broadly divided into two classes in terms of how they incorporate symmetry constraints. Invariant GNNs use scalar features such as bond distances and angles to describe the structure, ensuring that the predicted properties remain unchanged with respect to translation, rotation, and permutation. Equivariant GNNs, on the other hand, go one step further by ensuring that the transformation of tensorial properties, such as forces, dipole moments, etc. with respect to rotations are properly handled, thereby allowing the use of directional information extracted from relative bond vectors. For a comprehensive overview of different GNN architectures and their applications, readers are referred to recent literature Han et al. (2024) ; Duval et al. (2023) . Given sufficient training data, GNN architectures such as Nequip Batzner et al. (2022) , MACE Batatia et al. (2022) , Equiformer Liao and Smidt (2022) and many others Wang et al. (2024) ; Frank et al. (2024) ; Gasteiger et al. (2021) have been shown to provide state-of-the-art accuracies in the prediction of various properties and PESs. Chen et al. (2019) ; Fung et al. (2021) ; Bandi et al. (2024) ; Fu et al. (2022) Furthermore, unlike other MLIP architectures based on local-environment descriptors, GNNs have