Developing Autoencoder Models for Multi-Diagnostic Electron Density Processing in Tokamaks within a Scientific Machine Learning Framework

Abstract

Accurate reconstruction of plasma measurements is crucial for real-time control in magnetic confinement fusion devices. Diagnostic systems in these devices provide essential insights for physics studies and operational safety, yet they face a harsh environment where strong magnetic fields, radiation and neutron flux often degrade data quality. Noise, outliers and sensor faults make real-time processing a major challenge, traditionally requiring ad-hoc algorithms for each diagnostic. In this work, we present a multi-diagnostic integration approach based on autoencoders within a Scientific Machine Learning framework, used to process electron density measurements, as one of the most critical parameters to maintain under control during operations. By combining complementary measurements, such as Thomson Scattering, interferometry and polarimetry measurements, our method learns correlations across diagnostics, improving accuracy and resilience to faulty data. We investigate unsupervised and supervised autoencoder configurations, using synthetic tokamak scenarios generated in a virtual environment. An increasing level of multi-diagnostic integration shows a reduction in reconstruction errors, and, beyond accuracy, the proposed methodology enables fast inference suitable for real-time control. This approach enables robust reconstruction of key plasma parameters, supporting advanced control strategies in present tokamaks and informing the design of intelligent diagnostic systems for next-generation devices.