Cognitive systems build understanding of the world by making decisions based on past experiences. Scenarios which are
very different from those encountered in the past may lead to errors but provide data to enrich the knowledge base of
the system. Such systems should progressively improve as the variety of situations increases. Recent developments in
artificial intelligence enabled to quantify these vague notions of “understanding” or “experience” and made them mathematically
Xaedra employs methods of machine learning to enable selection of candidate materials based on a simple search for materials
properties specific to the application at hand. Of course, existence of data is a prerequisite for any machine learning
problem, it should be available, reliable, and curated, or effort has to be put in to create that data.
The Xaedra platform consists of a database containing known inorganic chemical compounds, their composition, and associated
crystallographic information. However, that experimental identification and verification of compound crystal structure
does not provide any information on the specific properties it might have. Some of the available materials property
data can be found in various forms ranging from peer-review journals, to material property databases. However, the information
is always very sparse. Xaedra gets around lack of empirical property data by using atomic level information and a neural
network to quantitatively predict previously unknown material properties. Thus, making it possible to do a properties-based
search for materials, even where no empirical property data exists.
At the heart of Xaedra is an atomic structure descriptor – a data structure uniquely identifying a material by associating
it with an atomic level fingerprint. The descriptor enables application of convolutional neural network and ultimately
construction of a regression model that can be trained to establish a mapping between a compounds atomic structures
and intensive properties of corresponding materials. This surrogate model is used to predict property values for all
known materials which are then stored in the database empowering the simple search. This, finally opens a path to rational
material design – a design-driven approach to materials exploration and discovery.
What academia is saying about materials and AI