Data-Driven estimation of experimental signals in TJ-II: A proof of principle

Magnetic confinement fusion experiments operate with extensive diagnostic coverage and high per-shot costs. Although machine operation is guided primarily by expert knowledge, theoretical understanding and decades of experimental experience, many aspects of scenario development still require exploring actuator trajectories whose consequences are uncertain until a discharge is executed. In this context, it is appealing to develop fast surrogate models capable of forecasting selected diagnostic signals, enabling in silico evaluation of candidate scenarios before committing machine time. Such capabilities would contribute to the long-term vision of digital twin infrastructures, which are beginning to gain traction in fusion but are not yet standard tools. As a proof of principle, we forecast the time evolution of the ACTON275 1 carbon 4 + (CV) line-emission signal in the TJ-II stellarator. The diagnostic consists of a 275 mm focal-length spectrometer tuned to λ = 227 . 1 nm; its signal correlates with line-averaged electron density, impurity sourcing and global radiation, making it a sensitive proxy for fuelling efficiency and impurity dynamics. Photons are converted into a voltage trace via a photomultiplier and amplifier stage and stored in the TJ-II database. We compare a classical baseline support vector regression (SVR) with three neural architectures: echo state networks (ESNs), long short-term memory networks (LSTMs) and an autoencoder-equipped LSTM (AE–LSTM), for multivariate time series forecasting of ACTON275 under electron cyclotron resonance heating (ECRH) conditions. Model selection is formulated as a multi-objective search favouring high average R 2 on unseen shots, low dispersion of R 2 and compact architectures compatible with real-time deployment. The best models reproduce the build-up and decay of CV emission with useful accuracy across a substantial fraction of discharges, suggesting that data-driven surrogates can form a building block towards future digital twin components for stellarator operation.