Materials-similarity metrics for the AFLOW data repository.
Abstract
Materials similarity metrics will be implemented for the AFLOW materials data repository. Machine-learning algorithms, such as convolutional neural networks and variational autoencoders, will be used to quantify the similarity between different materials, based on quantities such as the Bader characterization of the charge density, the electronic band structure fingerprint analysis and the elastic tensors. This will be used to identify materials with potentially similar behavior to materials known to display desirable properties such as superconductivity, superhardness, and topologically protected conducting states. Machine learning models will also be trained to predict the formation energies of materials. The AFLOWrepository contains formation energies for over 1.7 million materials, which makes for a large and diverse training set. The topological connectivity of materials structures will be represented graphically, and a convolutional neural network will be trained to identify trends and similarities between structures, and thus predict the formation energy of any structure prior to computationally expensive first principles calculations being performed. This will enable the estimation of the thermodynamic relevance of every possible decoration of each prototype in the AFLOW library, accounting for billions of potential materials. The entry pages for similar materials will be cross-linked with each other, allowing users to easily find materials with potentially similar behavior. A ???search for similar materials??? functionality will also be implemented, both as part of the aflow.org web portal and also within the AFLUX Search-API. A longer term goal of this research direction will be to implement a generative adversarial network to predict new structures. This will enable the generation of new entirely new materials structures that optimize targeted sets of properties, and that can act as new prototypes for enlarging the AFLOW database. The work proposed here will form an important foundation for this future direction.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 26, 2018
- Source ID
- N000141812613
Entities
People
- Stefano Curtarolo
Organizations
- Duke University
- Office of Naval Research
- United States Navy