Speaker
Description
High-bandwidth fluctuation diagnostics provide direct access to fast plasma dynamics but create a challenging real-time computing problem. As part of a U.S. DOE Genesis Mission project, we are developing an edge-AI architecture for streaming inference from megahertz plasma fluctuation measurements. We demonstrate this approach using beam emission spectroscopy (BES), which measures spatially resolved density fluctuations at 1 MHz, and investigate neural network designs suitable for FPGA-based processing of the continuous data stream. Starting from a convolutional network developed for turbulence-based plasma-instability-state prediction, we compare compressed architectures that differ in how temporal and spatial information are organized. Architectures preserving the temporal structure of the BES input retain the prediction performance of the baseline network while reducing the model to O(10^4) parameters. Quantization-aware training further converts the optimized network from 32-bit floating-point to 8-bit fixed-point representation with negligible performance degradation. FPGA synthesis indicates inference latencies of approximately 7–49 μs for the compressed architectures while remaining within device resource constraints. The operational relevance of real-time turbulence inference is further demonstrated on KSTAR, where neural-network-inferred turbulence information was incorporated into the plasma control system to trigger supersonic molecular beam injection. Together, these results establish a path from megahertz scientific data streams to FPGA-accelerated inference and turbulence-informed real-time control.
| Tutorial level (only for Tutorial) | Intermediate |
|---|