Gyrokinetic (GK) stability strongly influences the performance of high-confinement-mode pedestals in spherical tokamak plasmas. High-fidelity gyrokinetic codes such as GENE can model microinstability-driven transport, but the computational cost limits their routine use in integrated pedestal modeling workflows. Instead, present workflows often rely on reduced transport assumptions, such as the ballooning-critical pedestal model used in EPED. This work investigates machine-learning surrogate models for local linear gyrokinetic simulations in a MAST-U-relevant pedestal parameter space, with the aim of providing faster gyrokinetic-based inputs to reduced pedestal models. A sampling workflow is developed in which pedestal profile parameters are varied within experimentally motivated bounds and used to generate physically self-consistent Grad-Shafranov equilibria. This reduces the dimensionality of the data-generation problem compared with sampling local gyrokinetic inputs directly, while maintaining physically plausible combinations of plasma profiles, geometry, and local stability parameters. The surrogate models are trained to predict linear growth rates, real frequencies, and diffusivity-ratio transport fingerprints from local linear GENE simulations. A multi-head multilayer perceptron accurately reproduces the growth rate, while the diffusivity ratios and real frequency exhibit more clustered, regime-dependent behavior. A multi-head classification-regression model using frequency-based regime classes reduces the mean absolute error for these clustered targets and better captures sharp transitions associated with changes in the underlying instability regime, although errors near mode-transition regions remain a limitation.

In this work, we demonstrate the utility of state representation learning applied to modeling the time evolution of electron density and temperature profiles at ASDEX-Upgrade (AUG). The proposed model is a deep neural network which learns to map the high dimensional profile observations to a lower dimensional state. The mapped states, alongside the original profile's corresponding machine parameters are used to learn a forward model to propagate the state in time. We show that this approach is able to predict AUG discharges using only a selected set of machine parameters. The state is then further conditioned to encode information about the confinement regime, which yields a simple baseline linear classifier, while still retaining the information needed to predict the evolution of profiles. We then discuss the potential use cases and limitations of state representation learning algorithms applied to fusion devices.

Variational autoencoder (VAE)-based representation learning algorithms are explored for their capability to disentangle tokamak size dependence from other dependencies in a dataset of thousands of observed pedestal electron density and temperature profiles from JET and ASDEX Upgrade tokamaks. Representation learning aims to establish a useful representation that characterizes the dataset. In the context of magnetic confinement fusion devices, a useful representation could be considered to map the high-dimensional observations to a manifold that represents the actual degrees of freedom of the plasma scenario. A desired property for these representations is organization of the information into disentangled variables, enabling interpretation of the latent variables as representations of semantically meaningful characteristics of the data. The representation learning algorithms in this work are based on VAE that encodes the pedestal profile information into a reduced dimensionality latent space and learns to reconstruct the full profile information given the latent representation. Attaching an auxiliary regression objective for the machine control parameter configuration, broadly following the architecture of the domain invariant variational autoencoder (DIVA), the model learns to associate device control parameters with the latent representation. With this multimachine dataset, the representation does encode density scaling with device size that is qualitatively consistent with Greenwald density limit scaling. However, if the major radius of the device is given through a common regression objective with the other machine control parameters, the latent state of the representation struggles to clearly disentangle the device size from changes of the other machine control parameters. When separating the device size as an independent latent variable with dedicated regression objectives, similar to separation of domain and class labels in the original DIVA publication, the latent space becomes well organized as a function of the device size.

Within the 9th European Framework programme, since 2021 EUROfusion is operating five tokamaks under the auspices of a single Task Force called ‘Tokamak Exploitation’. The goal is to benefit from the complementary capabilities of each machine in a coordinated way and help in developing a scientific output scalable to future largre machines. The programme of this Task Force ensures that ASDEX Upgrade, MAST-U, TCV, WEST and JET (since 2022) work together to achieve the objectives of Missions 1 and 2 of the EUROfusion Roadmap: i) demonstrate plasma scenarios that increase the success margin of ITER and satisfy the requirements of DEMO and, ii) demonstrate an integrated approach that can handle the large power leaving ITER and DEMO plasmas. The Tokamak Exploitation task force has therefore organized experiments on these two missions with the goal to strengthen the physics and operational basis for the ITER baseline scenario and for exploiting the recent plasma exhaust enhancements in all four devices (PEX: Plasma EXhaust) for exploring the solution for handling heat and particle exhaust in ITER and develop the conceptual solutions for DEMO. The ITER Baseline scenario has been developed in a similar way in ASDEX Upgrade, TCV and JET. Key risks for ITER such as disruptions and run-aways have been also investigated in TCV, ASDEX Upgrade and JET. Experiments have explored successfully different divertor configurations (standard, super-X, snowflakes) in MAST-U and TCV and studied tungsten melting in WEST and ASDEX Upgrade. The input from the smaller devices to JET has also been proven successful to set-up novel control schemes on disruption avoidance and detachment.

A review is given on the highlights of a scatter-shot approach of developing machine-learning methods and artificial neural networks based fast predictors for the application to fusion exhaust. The aim is to enable and facilitate optimized and improved modeling allowing more flexible integration of physics models in the light of extrapolations towards future fusion devices. The project encompasses various research objectives: (a) developments of surrogate model predictors for power & particle exhaust in fusion power plants; (b) assessments of surrogate models for time-dependent phenomena in the plasma-edge; (c) feasibility studies of micro-macro model discovery for plasma-facing components surface morphology & durability; and (d) enhancements of pedestal models & databases through interpolators and generators exploiting uncertainty quantification. Presented results demonstrate useful applications for machine-learning and artificial intelligence in fusion exhaust modeling schemes, enabling an unprecedented combination of both fast and accurate simulation.

Predictive and real-time inference capability for the upstream separatrix electron density, ne, sep, is essential for design and control of core-edge integrated plasma scenarios. In this study, both supervised and semi -supervised machine learning algorithms are explored to establish direct mapping as well as indirect compressed representation of the pedestal profiles for predictions and inference of ne, sep. Based on the EUROfusion pedestal database for JET (Frassinetti et al., 2021), a tabular dataset was created, consisting of machine parameters, fraction of ELM cycle, high resolution Thomson scattering profiles of electron density and temperature, and ne, sep for 608 JET shots. Using the tabular dataset, the direct mapping approach provides a mapping of machine parameters and ELM percentage to ne, sep. Through representation learning, a compressed representation of the experimental pedestal electron density and temperature profiles is established. By conditioning the representation with machine control parameters, a probabilistic generative predictive model is established. For prediction, the machine parameters can be used to establish a conditional distribution of the compressed pedestal profiles, and the decoder that is trained as part of the algorithm can be used to decode the compressed representation back to full pedestal profiles. Although, in this work, a proof-of-principle for predicting and inferring ne, sep is given, such a representation learning can be used also for many other applications as the full pedestal profile is predicted. An implementation of this work can be found at https://github.com/ fusionby2030/psi_2022.

Pedestals are the key to conventional high performance plasma scenarios in tokamaks. However, high fidelity simulations of pedestal plasmas are extremely challenging due to the multiple physical processes and scales that are encompassed by tokamak pedestals. The leading paradigm for predicting the pedestal top pressure is encompassed by EPED-like models. However, EPED does not predict the pedestal top density, n(e,ped), but requires it as an input. EUROPED (Saarelma et al 2019 Phys. Plasmas 26 072501) employs simplified models, such as log-linear regression, to constrain n(e,ped) with tokamak machine control parameters in an EPED-like model. However, these simplified models for n(e,ped) often show disagreements with experimental observations and do not use all of the available numerical and categorical machine control information. In this work it is observed that using the same input parameters, decision tree ensembles and deep learning models improves the predictive quality of n(e,ped) by about 23% relative to that obtained with log-linear scaling laws, measured by root mean square error. Including all of the available tokamak machine control parameters, both numerical and categorical, leads to further improvement of about 13%. Finally, predictive quality was tested when including global normalized plasma pressure and effective charge state as inputs, as these parameters are known to impact pedestals. Surprisingly, these parameters lead to only a few percent further improvement of the predictive quality. The corresponding code for this analysis can be found at github.com/fusionby2030/supervised_learning_jetpdb.

Plasma-terminating disruptions in future fusion reactors may result in conversion of the initial current to a relativistic runaway electron beam. Validated predictive tools are required to optimize the scenarios and mitigation actuators to avoid the excessive damage that can be caused by such events. Many of the simulation tools applied in fusion energy research require the user to specify several input parameters that are not constrained by the available experimental information. Hence, a typical validation exercise requires multiparameter optimization to calibrate the uncertain input parameters for the best possible representation of the investigated physical system. The conventional approach, where an expert modeler conducts the parameter calibration based on domain knowledge, is prone to lead to an intractable validation challenge. For a typical simulation, conducting exhaustive multiparameter investigations manually to ensure a globally optimal solution and to rigorously quantify the uncertainties is an unattainable task, typically covered only partially and unsystematically. Bayesian inference algorithms offer a promising alternative approach that naturally includes uncertainty quantification and is less subjective to user bias in choosing the input parameters. The main challenge in using these methods is the computational cost of simulating enough samples to construct the posterior distributions for the uncertain input parameters. This challenge can be overcome by combining probabilistic surrogate modelling, such as Gaussian Process regression, with Bayesian optimization, which can reduce the number of required simulations by several orders of magnitude. Here, we implement this type of Bayesian optimization framework for a model for analysis of disruption runaway electrons, and explore for simulations of current quench in a JET plasma discharge with an argon induced disruption.