Minisymposium

Neural surrogate models of physics simulators are emerging ubiquitously in the Fusion community to satisfy the pressing need of fast optimisation tasks and flight simulator applications. However, gathering the training sets for these surrogates can be very expensive, and storing the data long-term may be impossible. In this talk I will demonstrate methodologies to obtain performing surrogate models at a significantly lower cost in terms of training data, with applications to 0D and 2D datasets and to a reactor-relevant streaming scenario.

Plasma physics simulations for fusion are particularly demanding in terms of computing time and memory consumption. Learning methods offer a way to reduce these costs, but at the expense of losing theoretical guarantees and explainability. To mitigate these shortcomings, we propose exploring hybrid numerical methods and models that combine classical approaches on the one hand and learning-based terms on the other. We apply these approaches to accelerate simulations of the Vlasov equation.We will present reduced modeling techniques for Vlasov-Poisson simulations in a weakly collisional regime, where we use an Euler equations solver coupled with a neural network model that describes the non-equilibrium part of the kinetic distribution. Through examples, we will demonstrate the neural network model's ability to capture the missing term while also discussing the limitations of these approaches.Finally, we will briefly introduce new ideas for hybrid neural network solvers and PIC codes to move toward more cost-effective simulations.

Recent advances in Artificial Intelligence (AI), Machine Learning (ML), and data-driven methods have opened promising pathways for addressing complex computational and control challenges in nuclear fusion and Tokamak research. Tokamak devices, a key configuration for magnetic confinement fusion, produce extensive and complex datasets that present unique challenges due to their multi-physics, nonlinear dynamics, and high-dimensional parameter spaces. This abstract introduces key computational challenges in fusion research, highlighting the critical importance of large-scale data analysis and advanced AI techniques for interpreting experimental results, optimizing plasma control strategies, and reducing operational uncertainties in contexts where high-bandwidth diagnostics produce large amounts of data, even during real-time operations. We will discuss how AI-driven surrogate and reduced-order models can effectively complement or even replace computationally expensive, high-fidelity simulation codes in certain scenarios, significantly accelerating analysis and enabling physics discovery. Another important aspect is the integration of real-time ML algorithms, transforming plasma control strategies and leading to significant performance improvements in extremely challenging fields such as plasma stability and disruption prediction. Specific examples and results from recent experimental and computational studies will be presented, demonstrating the efficacy and reliability of these data-driven approaches.

In this talk we perform structure-preserving reduced order modeling for the semi-discretized Hamiltonian PDEs. Reduced order modeling can alleviate the cost involved in applications such as optimization, uncertainty quantification and inverse problems, that require the repeated solution of large-dimensional physical systems. For this task we can use neural networks among other techniques.We start by giving a short overview of methods for performing reduced order modeling and how to make them structure-preserving. A focus will be put on symplectic autoencoders. These are neural networks that respect the symplectic structure of the underlying differential equation. We will show the advantages of both structure preservation and neural networks, compared to other techniques, when performing reduced order modeling for Hamiltonian systems and discuss future challenges ahead for dealing with real-world magnetic fusion plasmas.