MaterialsAtlas.org: A Materials Informatics Web App Platform for Materials Discovery and Survey of State-of-the-Art

Jianjun Hu, Stanislav Stefanov, Yuqi Song, Sadman Sadeed Omee, Steph-Yves M. Louis, Edirisuriya M. Dilanga Siriwardane, Yong ZhaoarXiv 2021 · arXiv

TL;DR

This paper introduces MaterialsAtlas.org, a user-friendly web platform designed to democratize materials discovery by providing easy access to a suite of computational tools. It addresses the current gap in accessible materials informatics web applications, aiming to accelerate research by enabling scientists without extensive computational expertise to explore, predict, and design new materials.

What Problem Does This Paper Address?

The field of materials informatics has seen rapid advancements in algorithms for materials property prediction, structure prediction, and generative design, largely driven by the availability of large experimental and computational datasets. However, a significant bottleneck exists: the lack of user-friendly web applications that make these powerful tools accessible to the broader materials science community. Many researchers, particularly experimentalists, lack the computational expertise to install, run, and adapt these complex algorithms locally. This underutilization of cutting-edge research hinders the pace of materials discovery. The paper highlights that while the bioinformatics field boasts thousands of web servers, materials informatics is still in its infancy, with a scarcity of diverse and high-quality web tools.

Methods

The core of this paper is the development and presentation of the MaterialsAtlas.org web platform. The platform is built using a Django backend, Vue.js frontend, Redis for job queuing, and Nginx as the web server, all containerized with Docker for easy deployment.

The platform offers four main categories of tools:

  • Composition and Structure Check and Validation: This includes tools for checking chemical validity (charge neutrality, electronegativity balance, Pauling rules) and thermodynamic stability (formation energy, e-above-hull energy). Machine learning models are employed for formation energy prediction, utilizing both composition-only (Roost algorithm) and structure-based (deep global attention graph neural networks) approaches.
  • Materials Property Prediction: This section leverages machine learning models to predict various material properties. Both composition-based (e.g., CrabNet) and structure-based (e.g., DeeperGATGNN) models are used. Predictions include 2D material formation, non-centrosymmetric materials, band gap, elastic moduli, hardness, thermal conductivity, ionic conductivity, and superconductivity.
  • Screening of Hypothetical Materials: The platform integrates generative models (like MATGAN and CubicGAN) to create databases of hypothetical materials. These databases can then be screened for novel compositions and structures, including cubic and 2D materials.
  • Utility Tools: A collection of helpful tools for materials scientists, such as a composition enumerator, feature generation pipeline, and tools for finding similar compositions/structures.

Datasets

The paper mentions the use of several publicly available datasets for training and evaluating their models:

  • Materials Project: A large database of computed materials properties, frequently used for training property prediction models (e.g., band gap, elastic moduli, non-centrosymmetric materials).
  • ICSD (Inorganic Crystal Structure Database): Used for training structure-based models and for thermal conductivity predictions.
  • 2DMatPedia: Used for training a classifier to predict 2D materials.
  • SuperCon database: Used for predicting superconductor transition temperatures.

The paper also implicitly refers to datasets generated by their own generative models (MATGAN, CubicGAN) for hypothetical materials.

Key Insights and Ideas

  • Democratizing Materials Informatics: The most significant idea is the creation of a unified, user-friendly platform to bridge the gap between advanced computational materials science and experimental researchers.
  • Addressing the "Web App Gap": The paper strongly advocates for the development of web applications as a crucial step for the widespread adoption and impact of materials informatics research, drawing parallels to the success of bioinformatics.
  • Integrated Workflow: MaterialsAtlas.org aims to support a significant portion of the materials discovery workflow, from initial validation and prediction to the generation and screening of novel candidates.
  • Leveraging Deep Learning for Prediction: The platform showcases the power of deep learning models (GNNs, transformers) for accurate materials property prediction, often outperforming older descriptor-based methods.
  • Generative Models for Novelty: The integration of generative adversarial networks (GANs) and variational autoencoders (VAEs) for creating hypothetical materials is a key aspect for exploring new chemical spaces.

Key Results

While specific quantitative results for each tool are detailed within the paper, the overarching outcome is the successful development and deployment of the MaterialsAtlas.org platform. The paper highlights:

  • High Accuracy in Predictions: Models achieve impressive accuracies, such as 88.98% for 2D material prediction and 84.8% for non-centrosymmetric material prediction.
  • User-Friendly Interface: The platform is designed with a focus on ease of use, allowing users to perform complex analyses without deep computational expertise.
  • Comprehensive Toolset: The platform offers a diverse range of tools covering multiple stages of the materials discovery process.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that some modules are still under development (e.g., e-above-hull energy, phonon check, ionic conductivity prediction due to limited data).
  • The performance of some models is noted as being for "experimental purposes" due to limited data size (e.g., thermal conductivity from ICSD).
  • The "over-smoothing" problem in graph neural networks is mentioned as a general challenge in the field, though their DeeperGATGNN model aims to mitigate this.
  • Future Work:
    • Phonon Calculation and Synthesizability Prediction: Developing ML-based classifiers for mechanical dynamic stability and synthesizability.
  • Crystal Structure Prediction: Investigating deep learning-guided algorithms inspired by protein structure prediction (AlphaFold).
  • Extensible Servers and API Services: Opening the platform to third-party web apps and providing REST API services for programmatic access.
  • Visualization and Interactive Exploration: Adding modules for visualizing materials property distributions in structural or composition space (e.g., using t-SNE).
  • Additional Utility Tools: Incorporating tools like structure file conversion and supercell generation.

Summary updated

BibTeX
@misc{hu2021materialsatlas,
  author = {Jianjun Hu and Stanislav Stefanov and Yuqi Song and Sadman Sadeed Omee and Steph-Yves M. Louis and Edirisuriya M. Dilanga Siriwardane and Yong Zhao},
  title = {{MaterialsAtlas.org: A Materials Informatics Web App Platform for Materials Discovery and Survey of State-of-the-Art}},
  year = {2021},
  eprint = {2109.04007},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.04007},
}
Figure