Speaker
Description
As part of the Coupling Experiment to Accelerator Controls (CEAC) project, an AI systems are being developed that adjust CEBAF beam and detector settings in response to experimental conditions measured in real time, with complementary use cases in MOLLER (HALL A) and GlueX (Hall D). For MOLLER, the goal is to minimize backgrounds in the Hall A Compton polarimeter, which must run continuously to reach the required precision on beam polarization, while preserving a sufficiently large spot size at the liquid hydrogen target and the quality of the beam modulation calibration. Uncertainty-aware AI surrogate models of the beamline, built from Elegant simulations and refined with CREX data, will guide a controller acting on the Hall A laser phase, beam position, and focusing magnets; with beam-jitter patterns serving as a real-time proxy for the target spot size; validation is planned with MOLLER commissioned data. For GlueX, the controller aims to to simultaneously minimize the uncertainties on the $\gamma p\to\rho X$ ($\rho\to\pi^+\pi^-$) and $\gamma p\to\eta X$ ($ \eta\to3\pi^0\to6\gamma$) cross sections for a fixed amount of beam time ($\sim$1 hour per decision). These channels have competing optimal beam currents because the reconstruction efficiency of the all-photon final state degrades faster with luminosity. The controller observes physics yields and uncertainties, detector rates and currents, DAQ livetime, and accelerator state. Its actions on beam current, drift chamber and TOF high voltages, tagger thresholds, and trigger settings are bounded by detector safety constraints. Analyses of archival GlueX data show that useful uncertainties can be extracted on this timescale, and Monte Carlo studies with overlaid random background quantify how reconstruction efficiency for both channels falls with beam current. The program culminates in a demonstration with dedicated beam time.
Meeting the hour-scale latency in GlueX requires AI-accelerated processing at each stage. A graph neural network that identifies correctly paired photons in $\eta\to3\pi^0$ candidates strongly suppresses split-off shower contamination and substantially improves the signal-to-background ratio over a multilayer perceptron baseline. A simplified fast-tracking mode increases the reconstruction rate several-fold with little loss of $\rho$ candidates, and neural-network tagger timewalk calibration is under validation. Data aggregation, online drift chamber calibration, and EPICS control run on SMOCS, a Kafka-based streaming platform, and reinforcement learning agents are developed with the Scientific Optimization and Controls Toolkit. We are also exploring teams of LLM-based agents for routine Hall D operations, evaluated against benchmarks such as current changes, current scans, and harp scans in TwinD, a sandbox simulation of the Hall D online system. These tools offer a template for autonomous, physics-aware operation of ePIC at the EIC.