AI tuning system reports 86.80% beam extraction rate on fusion ion sources

Researchers report in a preprint that an AI system achieved a beam-extraction rate of 86.80% across 409 shots while tuning five ion sources under supervised operation. The work addresses scientific instruments whose responses change as operating conditions evolve between shots. The system extracts physical measurements from waveforms, retrieves earlier operating experience and uses a multimodal language model to suggest adjustments. A separate feasibility check enforces constraints before execution. In a chronological replay of 455 shots, retrieved context increased mean agreement with historical adjustment directions by 22.4-42.7 percentage points. The physical demonstration establishes supervised tuning feasibility, rather than autonomous operation.