Machine Learning Data Platform

Machine learning and artificial intelligence techniques offer promising solutions in applications to modeling, diagnostics, control, and optimization of large particle accelerator systems and other large experimental physics facilities. Many research institutions have dedicated groups investigating machine learning and artificial intelligence applications for the accelerator sciences, such as the SLAC facility in Menlo Park, CA. To support these applications Osprey DCS is now building a “data-science ready” platform, the Machine Learning Data Platform, for developing and deploying machine-learning and general data science applications at such facilities. The platform has three primary functions: 1) high-speed data acquisition, 2) archiving and management of correlated, heterogeneous data, and 3) comprehensive, data-centric search and query capabilities of the archive. Additionally, data availability is rapid enough to support real-time applications for tuning online facility operations, and the archive maintains full data provenance as well as user notes, associations, and other annotations. A working platform prototype was developed in Phase I. The prototype may be deployed at any EPICSbased facility; however, the archiving system is now fully independent and standalone, any facility may use the data archiving, management, and search capabilities of the platform. Novel aspects include portability, comprehensive access to both data and metadata, and a standardized interface for machine-learning application development and deployment. Additionally, a standalone Web Application was developed, providing remote access with the archive using only a standard internet web browser. Phase II concerns performance improvements, advanced data science applications, and commercialization. The effort contains four primary focus efforts: 1) performance at a stated goal of acquisition and archiving of 4,000 signals at 1 kHz, 2) full archive user annotations including data calculations made post-acquisition (i.e., made from archive data), 3) fast first-level processing directly within the ingestion stream, 4) and direct algorithmic support through modular “plugins” to the platform. Phase II also includes platform installation and operation at existing accelerator facilities, the SLAC National Accelerator Laboratory and at Brookhaven National Laboratory. As a project milestone Osprey DCS will develop machine-learning based accelerator applications for platform deployment at these facilities, demonstrating platform installation, operation, and application. At Phase II completion, commercialization includes a packaged deployment system for facility installation and complete documentation for the platform; it will be available as an opensource system, complementing the EPICS control system popular at large-scale accelerator and experimental physics facilities.