Controlling a particle accelerator depends on models that describe how the beam responds to every adjustment of the magnets and cavities. Until now those models were built by hand and recalibrated periodically, a slow process that does not always keep pace with machine drift. The news presents digital twins that can generate themselves automatically from operating data and close the control loop: the model updates on its own and the tuning system acts on the machine in real time.

The underlying idea is that the digital twin should not be a static copy of the accelerator, but a model that continuously learns from observed behaviour. By integrating automatic model generation with the control layer, manual recalibration is removed and the gap between model and real machine is reduced. This is especially relevant in facilities where the beam must remain stable for hours of operation and where small deviations in magnetic fields or cavity phase degrade performance.

The approach has been deployed on an operating accelerator, where the auto-generated twin has been tasked with beam tuning without direct human supervision. The system takes beam diagnostic readings, reconstructs the model of the moment and proposes corrections that are applied to the control elements. Validation is done by comparing the predicted response with the measured one, so that the model corrects itself when it stops describing the machine well.

Closing the loop in this way opens the door to self-regulating accelerators that make better use of available beam time. It also raises questions about the reliability of learned models in operating conditions and about how to ensure that automatic corrections do not push the machine outside its safety margins. The next phase involves extending the deployment to more sections of the accelerator and studying the system's behaviour under more demanding operating conditions.