Speaker
Description
Digital twins and learned controllers are becoming central to accelerator operations, but a twin is only useful while it matches the machine, and few ML tools have reached sustained use. Beam settings change from run to run, and hardware ages in ways the diagnostics do not capture. These changes shift the relationship between machine inputs and beam response, so models trained on a fixed dataset drift away from the accelerator, forcing slow retraining and increasing operational downtime. Conditional models help with known setting changes but cannot cover every future configuration or any unmeasured drift. We discuss how lifelong (continual) learning (CL), which lets a model adapt to new machine conditions without forgetting earlier ones, can keep digital twins faithful over the life of the machine, and how it fits alongside CL for machine protection and anomaly detection and for tuning and control. We first demonstrate the prognostication use-case at the Spallation Neutron Source (SNS), where a conditional autoencoder for errant-beam anomaly detection was updated across 31 beam-setting configurations from real operational data. We then apply it to a digital twin of the Linac Coherent Light Source (LCLS) at SLAC, built as an adaptive surrogate of FEL pulse intensity that tracks the drifting machine data stream. A continual reinforcement learning (CRL) policy for tuning and active control is adapted on this continually updated twin, so the policy can keep improving without retraining from scratch or using beam time for exploration. We will present the results on measured data with offline setup mimicking online scenarios and our plan for online tests. Together, these results lay the groundwork for lifelong digital twins and controllers, and we describe our plan for testing them during machine operation at SNS and LCLS.