AI packs detector detail into a rapid particle trigger

An FPGA, or programmable-logic chip, processed a KOTO detector event in 20.7 microseconds. The 27 October preprint’s standalone demonstrator includes image assembly and neural inference; its projected 25.9-microsecond total includes unbuilt external transfers. The board replayed stored events. The projection fits a 32-microsecond budget.[1]

The image records energy deposited in a particle detector. KOTO’s cesium-iodide crystals convert incoming energy into light, which electronics turn into signals. KEK describes a calorimeter assembled from two crystal sizes, recording position, timing and energy. Some crystals previously served in Fermilab’s KTeV experiment before being transferred to Japan. Their physical measurements become a spatial pattern that a classifier can examine.[2]

The rare kaon-decay target includes an electron and positron; the common background has charged pions. The classifier rejected 99.34% of 50,000 recorded background events at nominal 95% retention of 50,000 simulated signal events; the rejection’s standard error was 0.04 percentage points.[1]

The design folds a 72-by-72 image into 18-by-18 locations with 16 channels. Fewer positions shorten serial processing, while extra channels permit parallel work.[1]

That rearrangement, called space-to-depth, changes the shape of the calculation without deleting measurements. TensorFlow’s documentation describes moving the values inside each small square into separate channels. A four-by-four patch becomes a stack of 16 values at one location. Their order still encodes their positions in the original patch.[3]

Replacing a patch with its average would erase the distinction between individual bright and dim cells. Rearrangement preserves those values: fewer spatial locations hold more channels at each location. A neural network can therefore receive the same underlying pattern in a different layout, with the local detail still available.[3]

Report an error or suggest a correction