Human–AI Collaboration: AI Agents Integrated into Fusion R&D Workflows, Advancing Diagnostic Signal Anomaly Detection
The Startorus Fusion R&D team integrated AI Agent–driven collaboration into the diagnostic signal processing workflow. Through efficient collaboration between carbon-based and silicon-based employees, the team replaced the conventional “train-from-scratch” neural network development paradigm. For three key tasks—magnetic probe signal reconstruction, AXUV anomaly detection, and IDS anomaly detection—development was completed in about one week (34 hours), compared with a previous cycle of roughly 200 hours, resulting in an overall efficiency gain of approximately 6×. new_line In terms of performance, magnetic probe signal reconstruction achieved an R² of approximately 0.97, delivering results broadly on par with conventional neural network methods. AXUV anomaly detection accuracy improved from roughly 86% to 98%, while IDS anomaly detection reached near-perfect accuracy, approaching 100%. new_line By combining physically grounded criteria and domain workflows defined by engineers with AI Agent–driven code generation and iterative optimization, this approach establishes a new R&D paradigm focused on rapid validation and model interpretability. Without requiring large-scale historical training datasets, it significantly accelerates diagnostic signal processing and has already been deployed in operational scenarios, including AXUV anomaly detection. The results validate the feasibility of domain expert + AI Agent collaborative development for fusion diagnostic signal processing, demonstrating tangible engineering and deployment value. new_line The following three comparison figures offer an intuitive view of how the two approaches differ in their outputs when applied to the same experimental dataset. new_line Figure 1 presents two experimental signals as examples to compare the two approaches. The upper subplot corresponds to the normal signal of Channel 7 in shot 260207066, while the lower subplot shows the anomalous signal of Channel 13 from the same shot. The light red shaded regions indicate intervals classified as anomalous, the blue curves represent the neural network model results, and the orange curves represent the AI Agent results. The comparison shows that, in the AXUV anomaly detection task, the AI Agent approach produces results highly consistent with those of the neural network model, accurately identifying both normal and anomalous channels. Moreover, it achieves higher overall detection accuracy, demonstrating strong effectiveness and reliability.
Figure 2 shows signal distributions for Channel 3 and Channel 32 in shot 260317048, displayed in the upper and lower subplots, respectively. The red curves represent the AI Agent detection results, while the blue curves correspond to the neural network model outputs. In the IDS anomaly detection task, the AI Agent and neural network approaches exhibit a high degree of consistency, both accurately identifying anomalous features with detection accuracy approaching 100%. These results validate the stability and engineering feasibility of the AI Agent approach in IDS anomaly detection scenarios.
Figure 3 presents prediction results for Magnetic Probe No. 32 and No. 38 in shot 251203024, shown in the upper and lower subplots, respectively. The orange curves indicate the AI Agent reconstruction results, while the green curves represent the neural network model outputs. As shown in the figure, for both probes in shot 251203024, the two methods accurately recover the signal variation trends and produce highly consistent predictions, with R² values exceeding 95%. These results demonstrate the effectiveness of the AI Agent approach and show that, even without complex training procedures, it can achieve performance comparable to that of neural network models.
