Fpga-accelerated Machine Learning Enables Real-Time Beam Emission Spectroscopy Diagnostics for Tokamak Plasma Control

Controlling the incredibly complex behaviour of plasma remains a central challenge in the pursuit of fusion energy, and accurate prediction of instabilities is crucial for sustaining stable, high-performance operation. Abhilasha Dave, James Russell, Mudit Mishra, and colleagues demonstrate a significant advance in this area by integrating a machine learning system directly into the real-time diagnostic infrastructure of the DIII-D fusion experiment. Their innovative approach utilises field-programmable gate arrays (FPGAs) to accelerate the analysis of beam emission spectroscopy data, enabling ultra-low latency forecasting of disruptive events within the plasma. This system, built using the SLAC Neural Network Library, uniquely allows for dynamic updates to the machine learning model without requiring hardware modifications, paving the way for adaptive control strategies and continuous improvement in plasma confinement and stability, a crucial step towards realising practical fusion power.

Machine Learning Improves Tokamak Plasma Control

Scientists are advancing fusion energy research by integrating machine learning into the real-time control systems of tokamaks, specifically the DIII-D facility. The goal is to improve plasma control, particularly by avoiding disruptions and achieving advanced operating regimes. This requires overcoming challenges related to computational delays and ensuring system reliability. The team successfully deployed a system utilizing a Field-Programmable Gate Array, a specialized processor, to accelerate the machine learning model and reduce latency. The resulting system achieves microsecond-scale inference latency and successfully integrates into the DIII-D control system, demonstrating the ability to classify plasma regimes and detect the onset of disruptive events. Researchers verified the ability to dynamically reload the neural network’s parameters on the FPGA without requiring a complete hardware redesign, enabling adaptation to changing plasma conditions. Future work focuses on bypassing the current CPU-based data acquisition system to further reduce latency and enable true streaming inference. They also plan to explore adaptive control strategies and investigate the deployment of more complex neural network architectures. This research represents a significant step towards realizing the potential of data-driven control in fusion energy, highlighting the importance of hardware acceleration for achieving real-time control and improving disruption avoidance.

Real-time Plasma Monitoring with Scalable Hardware

Scientists engineered a real-time plasma monitoring system for the DIII-D tokamak, integrating high-bandwidth diagnostics with the Plasma Control System to enable advanced control strategies. The system, built around the Scalable Hardware I/O Execution Layer for Diagnostics, a modular software framework, facilitates deterministic, low-latency data acquisition and communication across diverse compute platforms. This architecture streams data from diagnostics, including Beam Emission Spectroscopy, to processing hardware such as Field Programmable Gate Arrays and Graphics Processing Units for feature extraction and inference. Real-time data acquisition relies on Analog Input Cards operating at 500kHz, capturing signals from 64 channels of Beam Emission Spectroscopy with high precision.

These digitizers, housed within PCIe-based acquisition servers, are synchronized using a specialized framework, ensuring precise timing across the system. The team achieved a 1MHz sampling rate by utilizing dual analog-to-digital converters, capturing rapid fluctuations in plasma behavior. This node, which also includes a powerful GPU, interfaces directly with the Plasma Control System, enabling rapid actuator responses within microsecond-level latencies.

A key innovation is the library’s support for dynamic reloading of neural network parameters, allowing task-specific settings to be loaded at runtime without requiring full FPGA redesign. This flexibility enables multiple inference tasks, such as predicting instabilities and recognizing confinement regimes, to run on a single firmware-based model. The system achieves 4. 4 microsecond scale latency for neural network inference, demonstrating the feasibility of embedding dynamically reconfigurable hardware-accelerated machine learning into real-time fusion diagnostic pipelines.

Real-Time Plasma Control via Machine Learning

Scientists have achieved a breakthrough in real-time control of tokamak plasmas, essential for sustaining high-performance operation in future fusion reactors. The work demonstrates a hardware-accelerated machine learning system integrated into the DIII-D tokamak’s real-time diagnostic and control infrastructure, enabling ultra-low latency plasma state classification and edge-localized mode (ELM) forecasting. The system utilizes a Xilinx FPGA to process data from Beam Emission Spectroscopy, a diagnostic providing high-resolution measurements of electron density fluctuations. Experiments reveal the system achieves a neural network inference latency of 4.

4 microseconds, a critical advancement for responsive plasma control. This capability supports multiple classification tasks and adaptive control strategies, responding to evolving plasma conditions during live operation. Researchers successfully deployed a single firmware-based neural network model capable of both ELM forecasting and confinement regime recognition. The results establish a scalable and resilient path toward intelligent and autonomous plasma control, an essential capability for achieving reactor-relevant operation in future magnetic confinement fusion devices. By deploying a fully connected feedforward model on a dedicated FPGA, researchers achieved microsecond-scale inference, enabling accurate classification of plasma confinement regimes and early detection of edge-localized mode (ELM) events using data from Beam Emission Spectroscopy. The system interfaces seamlessly with existing control frameworks, supporting rapid decision-making crucial for disruption avoidance and maintaining stable plasma operation. The key achievement lies in the system’s ability to dynamically reconfigure the FPGA-based inference engine, allowing for updates to the neural network without requiring hardware resynthesis.

This adaptability supports continuous model refinement and facilitates seamless task switching during live operation, paving the way for more responsive and intelligent plasma control strategies. While the current data acquisition pipeline introduces some latency, the architecture demonstrates strong scalability for future deployments with more direct data interfaces. Future work will focus on establishing direct connections between data acquisition systems and the hardware inference engine, further reducing latency and enabling true streaming inference. This tighter integration will support more adaptive control strategies, critical for optimizing performance and preventing disruptions in next-generation tokamaks.