Ensuring reproducibility in the application of artificial intelligence to materials R&D

The Center for Hierarchical Materials Design (CHiMaD), a NIST Center of Excellence, established the Materials Data Facility (MDF) to build and maintain the data infrastructure for CHiMaD. In 2018, the MDF received support from the Department of Energy to create the Data and Learning Hub for Science (DLHub).

DLHub leverages existing CHiMaD data publication and sharing tools and builds a new set of capabilities to collect, publish, and categorize artificial intelligence models and other functions to be applied to data. DLHub allows others to run those models on leadership computing resources or in the cloud with a few lines of code, search among published models, and overall increase the reproducibility of machine learning applications to materials R&D (and beyond to the entire portfolio of scientific research). Example applications in X-ray science, batteries, and microscopy have been demonstrated.

The MDF/DLHub effort has now been further augmented through an National Science Foundation-supported collaborative Cyberinfrastructure for Sustained Scientific Innovation (CSSI) award to the University of Wisconsin-Madison and the University of Chicago that will smooth the interfaces between data providers (i.e., MDF) and AI-model providers (i.e., DLHub).

This effort seeks to collect and describe all of the best machine learning input datasets for materials and publish them in a well-structured format in the MDF. Concurrently, the DLHub will capture and describe the best AI models and build an interface to allow researchers to quickly pull data from MDF, including any number of well-aligned datasets, and execute the appropriate AI methods on DLHub for improved, reproducible, materials property prediction.