Chemistream: Big Data Materials on HPC Clouds

High performance computing (HPC) plays a key role in materials science, climate research, en- ergy technology and others. Artificial intelligence (AI) and machine learning (ML) algorithms are becoming a way of rapidly predicting materials properties. However, recent surveys have shown an under-representation by companies that could leverage HPC technologies and AI/ML meth- ods. One reason is that codes of interest can be complex to build and use and a second reason is that moving away from local workstations to take advantage of HPC resources can be filled with difficulties particularly for small and medium commercial users. We will extend the capabilities of the ’Chemistream’ framework in order to provide: access to a va- riety of material properties databases (some funded by DOE), access to specific AI/ML algorithms and advanced workflows to optimize molecular design using AI/ML models. During Phase I we integrated the use of S3 persistent storage at Amazon Web Services (AWS) and Azure ’Blob Storage’ into the Chemistream application. During Phase I we identified the steps needed to extend the NREL Bond Dissociation Energy (BDE) database. These steps include accessing persistent cloud storage for updating current database information and running DFT calculations using NWChem. We also identified the steps for using AI/ML techniques to predict partial charges using the NREL organic photovoltaic (OPV) database. Lastly, this work has been prototyped in examples integrated into Chemistream and running on cloud resources. The proposed Phase II work will develop integrated access to AI/ML molecular databases for training new models. This unified interface will enable access to pre-existing molecular databases as well as custom workflows for generating specific, user-defined training data. Phase II work will also streamline development for AI/ML training models to predict small molecule properties relevant to the nanotech and pharmaceuticals industries. Finally, Phase II work will prototype examples for optimizing the design of advanced materials with Reinforcement-Learning (RL). This progress will positively impact one of the goals of the Materials Genome initiative to enhance the rate of breakthroughs in complex materials chemistry and materials design. Leveraging these advances in computational chemistry, artificial intelligence, machine learning and materials re- search will allow small/medium companies to more effectively compete against larger companies.