25–29 May 2026
La Biodola - Isola d'Elba (Italy)
Europe/Rome timezone
NB: The submission deadline for the Student Paper Awards is Monday, 11 May.

27 FPGA-Deployed Variational Autoencoder for Real-Time Soft X-Ray Electron Temperature Reconstruction in RFX-mod2

27 May 2026, 10:20
2m
Maria Luisa Room (Hotel Hermitage)

Maria Luisa Room

Hotel Hermitage

Mini Oral AI, Machine Learning, Real Time Simulation, Intelligent Signal Processing Mini Orals

Speaker

Luca Orlandi (Università di Padova - Consorzio RFX)

Description

RFX-mod2 is the upgraded version of the RFX-mod Reversed-Field Pinch (RFP) device, operated at Consorzio RFX. Given the substantial phenomenological and engineering complexity of the device, control performance is expected to benefit significantly from the integration of a broader set of diagnostics, such as Soft X-Ray (SXR) measurements, that help characterize the internal plasma state. A crucial parameter for accurately assessing this state is the electron temperature, inferred from SXR radiation. However, to effectively integrate this information in real time, it is necessary to derive a denoised, low-dimensional representation of the acquired signal within a timeframe compatible with the control cycle.
Within this context, autoencoder (AE) architectures and their variational extension (VAE) can be employed to capture the strong non-linearities characteristic of plasma behavior. The proposed VAE model, initially developed in earlier work at the RFX Consortium and subsequently extended, has been evaluated on two datasets: one synthetic and one experimental, the latter derived from the SXR diagnostic of RFX-mod. To achieve real-time deployment on FPGA hardware, the model is subjected to a quantization procedure using either the CERN/HLS4ML framework or the Xilinx/FINN framework. The system demonstrates the ability to recover relevant physical information in real time, suitable for integration into plasma discharge control along-side other diagnostic systems. Results show that the VAE can consistently reconstruct the input temperature profile, even under moderate to high levels of data loss, highlighting the potential of deep generative models for robust data imputation in fusion plasma diagnostics.

Minioral Yes
IEEE Member No
Are you a student? Yes

Authors

Andrea Rigoni Garola (RFX) Prof. Lidia Piron (Università di Padova) Mr Lorenzo Saccaro (Università di Padova - Consorzio RFX) Luca Orlandi (Università di Padova - Consorzio RFX) Dr Marco Gobbin (ISTP - CNR) Dr Paolo Franz (Consorzio RFX) Dr Roberto Cavazzana

Presentation materials

There are no materials yet.