Researchers have detailed a new end-to-end workflow for deploying recurrent probabilistic neural networks (RPNNs) on field-programmable gate array (FPGA) hardware for real-time plasma state estimation. According to a preprint released on arXiv, the method achieves a deterministic single-timestep inference latency below 10 microseconds. This speed meets the demanding cycle-time requirements for integration into a tokamak's plasma control system (PCS), specifically for use within a model-predictive control architecture. The work addresses the significant computational challenge of providing fast, accurate plasma state data, which is critical for preventing disruptions and maintaining stable operation. Source: arXiv
The approach combines several optimization techniques to achieve its performance on a Xilinx Alveo U50 device. The core strategies include reducing the neural network's architecture size and implementing quantization-aware training using the QKeras library. Quantization reduces the numerical precision of the model's weights and activations, which significantly decreases computational load and memory footprint without an unacceptable loss of accuracy. The trained and quantized model is then synthesized for the specific FPGA hardware using the hls4ml toolchain, a process that translates high-level machine learning models into low-level hardware descriptions. This workflow is designed to be a comprehensive solution from software model to hardware deployment. Source: arXiv
The approach combines several optimization techniques to achieve its performance on a Xilinx Alveo U50 device.
Effective plasma control is fundamental to achieving high-performance, steady-state operation in magnetic confinement devices. Control systems must react to rapid changes in plasma parameters like position, shape, and temperature to sustain the reaction and avoid instabilities that can lead to a disruption. Traditional control methods often rely on simplified physics models that may not capture the full complexity of plasma dynamics. The use of sophisticated, data-driven models like RPNNs offers a path to more accurate state estimation, but their computational intensity has historically made them too slow for the microsecond-scale feedback loops required in a live tokamak environment. Source: arXiv
The reported design successfully fits within all four primary resource budgets of the target FPGA: digital signal processors (DSP), look-up tables (LUT), flip-flops (FF), and block random-access memory (BRAM). This confirms the hardware viability of the synthesized model, indicating it does not require over-provisioned or exotic hardware to run. By demonstrating a concrete method for accelerating complex, probabilistic models on accessible hardware, this work provides a template for developing next-generation control and diagnostic tools. Such tools are essential for the operational success of future devices, including commercial-scale power plants where high availability and disruption avoidance will be paramount. Source: arXiv