Berk Onat (Engineering)
Structural Representations of Materials for Machine Learning using the Novel Materials Discovery Big-Data Analytics Platform
Data-driven analysis has received increasing attraction as the need for extracting and classifying information becomes daunting as a result of the large amount of data that is now available. While it is becoming common practice in many research fields, big-data analytics is still under-exploited in material science and engineering. A European project, NOMAD, the Novel Materials Discovery Laboratory addressed this problem of creating a common platform for big-data analytics through a large data base of computational materials science calculations. To achieve this goal, NOMAD is built on a repository of tools, codes and standardised data from many ab-initio and molecular dynamics codes in an open-access sharing environment. In this talk, I will present our development to core components of the NOMAD architecture and contributions to the code-independent format of the materials data in NOMAD archives. I will also present recent affords to analyze the performance of various representations for structural description of materials (e.g. Behler-Parrinello symmetry functions, SOAP) in machine learning on bulk and molecular datasets of the NOMAD archive.