Speaker
Description
A fundamental challenge in large-scale accelerator operation is to infer internal machine states from limited monitoring. In the J-PARC linac, errors in RF cavities lead to momentum drifts observable through beam phase monitors. Although the number and placement of monitors satisfy the solvability condition, the cumulative and sequentially coupled nonlinear nature of beam dynamics makes the inverse problem numerically ill-conditioned.
Our initial machine-learning study demonstrated that a multilayer perceptron can accurately model the forward mapping from cavity errors to observed phases. However, the inverse model, which predicts cavity errors from monitor signals, failed under the non-redundant condition where each cavity was followed by only one phase probe. In contrast, when at least two probes were available (redundant case), the inverse prediction became accurate. This brought to light a fundamental limitation: what appeared to be a dependence on monitor redundancy was, in fact, a consequence of the model’s limited capacity to represent cumulative dependencies and of the insufficient information content in the training data.
We reformulated the problem within an encoder/decoder framework, which treats the accelerator as a causal sequence of coupled elements. The inverse prediction succeeded in the non-redundant case. This reveals that physical irreversibility can be effectively mitigated when the model captures the hidden correlations underlying the sequential process.
| In which format do you inted to submit your paper? | Word |
|---|---|
| Preprint marking on your proceeding paper | I wish my paper to be marked as preprint. |