AFLOW-CHULL: Cloud-Oriented Platform for Autonomous Phase Stability Analysis

Corey Oses, Eric Gossett, David Hicks, Frisco Rose, Michael J. Mehl, Eric Perim, Ichiro Takeuchi, Stefano Sanvito, Matthias Scheffler, Yoav Lederer, Ohad Levy, Cormac Toher, Stefano CurtaroloJournal of Chemical Information and Modeling 2018 · OpenReview

TL;DR

This paper introduces AFLOW-CHULL, an open-source platform for automating the prediction of material phase stability. By leveraging a massive database of calculated material properties and advanced computational techniques, AFLOW-CHULL can rapidly identify promising new materials for synthesis, significantly accelerating the discovery process.

What Problem Does This Paper Address?

Discovering new functional materials is a complex endeavor, hampered by the sheer number of possible chemical combinations and the difficulty in predicting their stability. While computational methods like density functional theory (DFT) can predict material properties, determining which of the countless hypothetical structures are actually stable under various conditions is a major bottleneck. Existing methods often require extensive manual effort and specialized knowledge, limiting their widespread application. This paper addresses the need for an automated, accessible, and robust platform to perform these crucial thermodynamic stability analyses, bridging the gap between theoretical prediction and experimental realization.

Methods

AFLOW-CHULL builds upon the existing AFLOW (Automatic Flow) framework and integrates with the AFLUX Search-API to access a vast repository of over 1.8 million calculated material properties. The core of the module is its ability to construct N-dimensional convex hulls, a standard method for determining thermodynamic stability. The algorithm is inspired by the Qhull algorithm but adapted for thermodynamic hulls, handling the unique challenges of stoichiometric and energetic coordinates.

Key methodological innovations include:

  • Dimensionally-Iterative Hull Construction: To address issues with lower-dimensional data in higher-dimensional hull constructions (e.g., vertical or content-less facets), AFLOW-CHULL iteratively calculates hulls, starting with binary systems and propagating results to ternary and higher-order systems.
  • Outlier Detection: A robust outlier detection algorithm is employed to discard potentially erroneous data points that could distort the convex hull and lead to incorrect stability predictions.
  • Duplicate Detection: A procedure using AFLOW-XTAL-MATCH identifies and removes structurally equivalent entries, ensuring the integrity of the dataset.
  • Thermodynamic Descriptors: Beyond simple stability, AFLOW-CHULL calculates various descriptors, including the energy driving decomposition (ΔHf) for unstable phases and a stability criterion (dsc) for stable phases, which quantifies their robustness.
  • Balanced Decomposition Reactions: The module can determine the full chemical equation for the decomposition of unstable phases into stable ones.

Datasets

The primary dataset utilized is the AFLOW.org repository, which contains data for over 1.8 million compounds and 180 million calculated properties. This data is generated using DFT calculations performed with VASP, adhering to the AFLOW Standard for high-throughput materials science. The paper also references the ICSD (Inorganic Crystal Structure Database) and the ASM Alloy Phase Diagram database for comparison and validation.

Key Insights and Ideas

  • Automation is Key: The development of AFLOW-CHULL demonstrates that automating complex thermodynamic analyses can unlock the potential of large materials databases for rapid discovery.
  • The Power of the Convex Hull: The convex hull remains a fundamental and powerful tool for understanding phase stability, and its application to vast datasets is crucial for materials discovery.
  • Beyond Stability: Robustness Matters: Identifying not just stable phases but also their robustness (using the dsc criterion) is vital for predicting materials that will remain stable under realistic conditions and for guiding synthesis.
  • Structure-Property Relationships: The prevalence of certain crystal structures (like F43m #216) among the most stable predicted ternary phases suggests underlying principles that warrant further investigation.

Key Results

AFLOW-CHULL was applied to the entire AFLOW.org repository, resulting in the thorough thermodynamic characterization of over 1,300 binary and ternary systems. This analysis led to the identification of several promising candidate phases for synthesis, including:

  • 18 new C156-type structures.
  • Two new half-Heusler compounds.

The paper also highlights the successful prediction of new magnets through computational approaches, demonstrating the practical impact of such tools. The web application developed alongside AFLOW-CHULL provides an interactive platform for exploring these results.

Limitations and Future Work

  • Data Convergence: While AFLOW-CHULL relies on converged DFT calculations, the paper acknowledges that there's never a guarantee that all stable structures have been identified, especially for complex systems.
  • Disorder: The current AFLOW.org repository excludes disordered (partially occupied) structures, which are important for certain material classes (e.g., high-entropy alloys). Future releases aim to integrate methods for handling disorder.
  • Experimental Validation: While the tool identifies promising candidates, experimental validation is still a necessary step. The paper suggests that AFLOW-CHULL can significantly narrow down the search space for experimentalists.
  • Repository Expansion: The authors suggest that AFLOW-CHULL can be easily extended to other materials repositories beyond AFLOW.org, broadening its applicability.
  • Larger Structure Prototypes: The prominence of certain space groups in ternary predictions might be influenced by sampling bias. Further exploration of larger structure prototypes and a wider range of space groups is needed.

Summary updated

BibTeX
@article{oses2018aflow,
  author = {Corey Oses and Eric Gossett and David Hicks and Frisco Rose and Michael J. Mehl and Eric Perim and Ichiro Takeuchi and Stefano Sanvito and Matthias Scheffler and Yoav Lederer and Ohad Levy and Cormac Toher and Stefano Curtarolo},
  title = {{AFLOW-CHULL: Cloud-Oriented Platform for Autonomous Phase Stability Analysis}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2018},
  url = {https://openreview.net/forum?id=QdIXEflESg},
}
Figure