In 2009 Dr. Coffee lead the first time-resolved experiment at SLAC's x-ray free electron laser. The upcoming TB/s data velocities for next generation of X-Ray lasers has led Dr. Coffee to move inference away from the data center and into the instruments and sensors themselves.
EdgeAI: Ultrafast inference in an age of ultrahigh data velocity
The next generation of X-ray Free Electron Lasers will produce basic science data at velocities that are expected to exceed TB/s. The scientifically relevant information that is contained in that data is much more human intuitive and theory compatible when represented as the physical observable under study.
These representations, or non-linear mappings, can be thought of as lossy compression algorithms whereby experimental particularities and noise are "lost" yet physically relevant information is preserved. We therefore seek a solution that capitalizes on very high throughput, ultra-low latency, statistical inference that preserve scientific transparency while achieving TB/s data ingestion rates at the site of production.
At x-ray free electron laser facilities, the meaning of data changes weekly; each fresh user group targets a different variety of physical observables but leverages the same detectors. Probing planetary science, artificial photosynthesis, superconductivity and spontaneous symmetry breaking in molecules, each group embodies a unique combination of human knowledge, coalescing around a semi-autonomous accelerator control system and a continuously re-purposed suite of detectors. Capable of x-ray pulses as short as the orbit time of an electron around the atomic nucleus, the commonly schizophrenic stochastic dance of multiple colors and polarizations forms a wild variability that requires individual treatment of each shot of the x-ray laser, up to a million shots per second for SLAC/Stanford's LCLS-II that must be processed in flight. The inference engine must be fully transparent to analytical investigation and must incorporate the collection of human expertise to "nudge" the automated control systems based on scientific "hunches" or we risk losing human creativity in the scientific pursuit. Our example of resonant x-ray spectroscopy will help to focus our attention on the unique needs in basic science for Edge AI assistance with on-the-fly data analysis and experimental control in discovery science.
