MaterialNet: A web-based graph explorer for materials science data

Roni Choudhury, Muratahan Aykol, Samuel Gratzl, Joseph H. Montoya, Jens S. HummelshøjJ. Open Source Softw 2020 · DOI · DBLP

Introduction

The advent of "Materials Genome" initiatives has led to the creation of massive databases containing properties for tens of thousands of materials. While these repositories are invaluable for specific lookups, they often lack the tools necessary for holistic exploration. Traditional spreadsheet-based views fail to capture the relational nature of materials science, where properties like phase stability and crystal structure similarity are inherently contextual. MaterialNet addresses this gap by providing an open-source, web-based platform that represents materials as nodes in a graph, allowing researchers to navigate complex datasets as interactive "maps."

Methodology

MaterialNet is engineered for high-performance visualization of large-scale relational data within a browser environment. Its architecture leverages several specialized libraries to handle the computational load of rendering thousands of data points:

  • GeoJS: Originally designed for geographic mapping, it is repurposed here to render interactive node-link diagrams, enabling the smooth handling of graphs with tens of thousands of nodes and hundreds of thousands of edges.
  • LineUp: This component provides an interactive tabular view, allowing users to rank and sort materials by multiple attributes (e.g., formation energy or year of discovery) while maintaining synchronization with the graph view.
  • VTK.js: Integrates a 3D viewer for inspecting the atomic crystal structures of individual materials directly within the interface.
  • Force-Directed Layouts: These algorithms are employed to visually cluster related materials, providing an intuitive spatial representation of chemical families and stability groups.

Datasets and Relational Context

The utility of MaterialNet is demonstrated through several large-scale networks that highlight different aspects of material relationships:

  1. Materials Stability Network: Maps thermodynamic relationships across approximately 20,000 materials and 200,000 links.
  2. MatScholar: A text co-occurrence network where links represent materials that frequently appear together in scientific literature.
  3. Materials Similarity Network: A graph where edges represent structural or chemical similarities between compounds.

By treating these databases as maps, researchers can leverage spatial intuition to identify "neighborhoods" of interest. This relational context is crucial for "predictive synthesis," helping scientists understand how a material's position relative to stable, known compounds might inform its experimental creation in a laboratory setting.

Results

The primary contribution is the MaterialNet tool itself (accessible at maps.matr.io), which demonstrates significant scalability by visualizing networks of 20,000 materials and 200,000 links with interactive performance.

The MaterialNet interface showing the neighborhood of the undiscovered material LiOsO3.

A key use case involves the exploration of "undiscovered" materials—compounds that are computationally predicted but not yet synthesized. As shown in the interface, a researcher can locate a hypothetical material like $LiOsO_3$ and examine its proximity to stable neighbors like $Li_2O$. This macro-to-micro workflow—moving from a cluster in the graph to a filtered list in the table, and finally to the atomic structure in the 3D viewer—streamlines the discovery process by allowing researchers to spot trends and gaps that would be invisible in a standard list.

Conclusion and Future Work

MaterialNet provides a powerful framework for navigating the "landscape" of materials data, moving beyond simple lookups to relational exploration. Future developments aim to enhance search capabilities through a domain-specific language (DSL) for complex queries and to introduce new visualization modes for specific chemical subspaces. The architecture remains extensible, with potential applications for experimental synthesis recipes and patent-based networks.

Summary updated

BibTeX
@article{choudhury2020materialnet,
  author = {Roni Choudhury and Muratahan Aykol and Samuel Gratzl and Joseph H. Montoya and Jens S. Hummelshøj},
  title = {{MaterialNet: A web-based graph explorer for materials science data}},
  journal = {J. Open Source Softw},
  year = {2020},
  doi = {10.21105/joss.02105},
}
Figure