Real-time data reduction at 100 Tbps: Challenge and opportunity for AI-based data reduction for next-generation large-scale nuclear physics collider experiment
The modern large-scale nuclear physics (NP) experiments in high-energy particle colliders utilize streaming-readout electronics to digitize detector response at O(100) Tbps bandwidth. Prominent examples at Brookhaven National Lab (BNL) include the sPHENIX experiment at Relativistic Heavy Ion Collider (RHIC), which is under construction, and the experiments proposed for the Electron-Ion Collider (EIC), planned for the 2030s . One of the main challenges for these streaming readout systems is to manage the data rate with sufficient data reduction in real time so the end-data fit persistent storage for offline analysis, which is typically at O(1000) times smaller and O(100) Gbps. Such data reduction traditionally is achieved via real-time high level triggers, which select and save a small subset of collisions of interest. Although triggering is applicable to high energy collider experiments such as those at the Large Hardron Collider at CERN, it is insufficient for these nuclear physics experiments which study diverse collision topologies. And traditional triggering approach is inefficient to preserve the max information harvested from the operation of colliders that costs O(100)M per year to DOE. Meanwhile, in recent years, ML-based high-throughput data reduction has emerged as a promising approach to efficiently preserve max information for a given space of persistent storage, e.g. via AI data compression, feature extraction, and noise filtering.