Advancing MARFE detection in JET’s operational camera videos through Machine Learning techniques - Article

Sep 17, 2026

Uploaded by Noah

To demonstrate the capability of operational cameras in nuclear fusion devices, videos from JET operational cameras were used to detect MARFEs, an edge plasma phenomenon. Three detection approaches were tested: two previously published methods and a new method based on intensity masks. After validation, the produced outputs enabled several Machine Learning models, culminating in a final model that combines camera data with signals and diagnostics from other JET instruments, achieving 96.9% accuracy.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

MARFE

Data and detection models

MARFE database construction

Existing-model adaptations

Machine Learning models

Model architectures

Results and discussion

Prediction analysis

Conclusions

Fusion Engineering and Design 205 (2024) 114534 | Advancing MARFE detection in JET’s operational camera videos through Machine Learning techniques
A. González Ganzábal a,b,∗, G.A. Rattá a, D. Gadariyaa, S. Dormido-Cantoc, JET Contributors 1 a Laboratorio Nacional de Fusión, Spain
b Universidad Nacional de Educación a Distancia, Spain
c Dpto. Informática y Automática, Universidad Nacional de Educación a Distancia, Spain | | | | | --- | --- | --- | --- | | A R T I C L E I N F O | | A B S T R A C T | | | Keywords:
JET
Machine learning MARFE Operational cameras Video detection
Visible cameras | | In order to prove the capability of operational cameras in nuclear fusion devices, the videos from the cameras at JET were used to detect the occurrence of MARFEs, an edge plasma phenomenon. Three techniques were tested in this work: two already reviewed in other publications and a new one based on intensity masks. Once these methods were validated, their output was used to develop several Machine Learning models to improve performance. A final Machine Learning model was devised using both data from the operational cameras and several signals and diagnostics from other instruments at JET. The outcomes achieved using all the methods presented were deemed satisfactory, leading to the final Machine Learning model exhibiting an impressive accuracy rate of 96.9%. Furthermore, the models allow for detection both in frame by frame (if only video data is used) and in 2 ms time steps should all diagnostics be used. | | 1. Introduction In the realm of energy prospects, nuclear fusion emerges as the flagship of clean energy for the future. With its promise of being a sustainable and nearly limitless resource, nuclear fusion holds tremendous potential. This process involves the fusion of light elements, primarily helium (He), deuterium (2 􀁈), and tritium (3 􀁈), resulting in the release of vast amounts of energy. To recreate the extreme conditions necessary for these reactions, plasma — an ionised gas generated by electrical currents—is employed in experimental nuclear fusion devices. Among these devices, JET stands as the largest operational tokamak of its generation, before ITER and JT-60SA, which has started its operations recently. To gain a better understanding of various physical phenomena and develop effective control mechanisms, multiple diagnostic tools have been implemented. This paper will specifically explore the significance of operational cameras as valuable diagnostic tools. Operational cameras possess substantial untapped informational value, despite being underutilised in comparison with other diagnosticsin several nuclear fusion devices. In this paper, the value of these operational cameras for plasma diagnostics will be proven by analysing the detection of MARFEs (Multifacetic Asymmetric Radiation From the Edge2 ), an edge plasma phenomenon with data taken directly from JET’s operational cameras. After a brief introduction regarding MARFEs as a physical phenomenon and JET’s operational cameras (Sections 1.1 and 1.2, respectively), this paper will delve into the construction process of the MARFE database in Section 2 and will examine various detection models developed exclusively for this work (see 2.1) or adapted from existing publications by other authors (2.2 and 2.3). Furthermore, models with Machine Learning architecture were also created and will be explored in Section 3. Subsequently, we present the derived results and conduct a comprehensive analysis of the predictions, drawing comparisons among them in Section 4. The final conclusions of this work can be found in Section 5. 1.1. MARFE MARFEs are low-temperature, high-density regions that appear atthe plasma periphery when the electron number density surpasses a critical threshold. Adjusting parameters such as the plasma current or reducing lighter impurities can increase this threshold as it was experimentally found [1]. Understanding the

FAQs

What is MARFE detection used for in this study?

MARFEs are edge plasma phenomena whose detection is crucial for avoiding and controlling disruptions that can severely damage nuclear devices.

Which techniques were tested for MARFE detection from operational-camera videos?

The study tested three techniques: two already reviewed methods and a new approach based on intensity masks.

How does the final Machine Learning model improve performance and what accuracy was achieved?

After validating the detection methods, the authors trained Machine Learning models using operational-camera outputs plus multiple signals and diagnostics from other JET instruments, reaching 96.9% accuracy.

Sep 17, 2026

Uploaded by Noah