Speaker
Description
This contribution presents initial efforts towards a machine-learning framework for estimating latent ECR source states from operational data collected at the Facility for Rare Isotope Beams (FRIB). The approach combines time-series measurements from source controls, beam transport elements, and current diagnostics to infer hidden variables associated with plasma conditions, ion production, and beam transport.
The work is motivated by the limited availability of direct plasma diagnostics during routine operation, and by an earlier study of a baseline predictive model, which showed that predictive accuracy alone does not guarantee physically meaningful internal representations. Building on this finding, we are designing a hierarchical latent-state architecture that mirrors the physical signal path from source controls through plasma formation, extraction, and transport, with each stage supervised, where possible, by intermediate diagnostic measurements, so that the learned representations remain tied to source physics rather than incidental correlations.
This work is an early step toward an operational digital twin of the FRIB ECR ion source. We present the motivating diagnostic findings, the proposed architecture, and open questions ahead of implementation, with an eye toward future applications such as fault diagnosis, performance prediction, and AI-assisted tuning.
Funding Agency
DOE, Office of Science, Office of Nuclear Physics, under Award No.DE-SC0024707
DOE Office of Science User Facility under Award No. DE-SC0023633
| Classification | MC8: Source modelling |
|---|---|
| I have read and accept the Privacy Policy Statement | Yes |