US researchers at the SLAC National Accelerator Laboratory have developed the Scalable Particle Imaging with Neural Embeddings (SPINE). It is a machine learning tool to analyze data from experiments designed to find neutrinos, the elusive ghost particle. The tool recently analyzed its first set of data from a real-world experiment and is prepared to tackle data from more ambitious projects in the future. 

Neutrinos are extremely abundant particles in our universe but are the most elusive. These nearly massless particles are released by stars and and radioactive elements alike but rarely interact with matter. Despite their abundance, these have been very difficult to detect. So, scientists have built large particle accelerators and complex detectors to catch these ghostly particles. 

A typical neutrino detector holds hundreds of tons of liquid argon cooled to temperatures of -300 degrees Fahrenheit (-184 degrees Celsius). When a neutrino strikes an argon atom, it sends several particles traveling through the liquid and leaves a trail of electrons that are picked up by collection wires to create a three-dimensional image of the event. 

Capturing the image is only the beginning, since scientists must reconstruct the event from a complex picture generated by data from the collection wires, overlapping signals, and background activity, which can include multiple neutrino interactions at the same time. 

Conventionally, scientists have done this using special algorithms that chain events together one step at a time. Each algorithm performs a narrow task and passes it to the next. However, this carries a significant risk: an error at any step can carry forward through the rest of the computation.

Source link