Direct-Modelling in Nuclear Fusion by Combining Data-Integration and Physics-Informed Neural Networks

Research on nuclear fusion aims to develop a carbon-free, sustainable, limitless, and safe energy source for both medium- and long-term applications. One of the possible ways to control nuclear fusion reactions is based on magnetic confinement. Tokamaks are nuclear fusion reactors that generate a toroidal field by external coils and poloidal magnetic field by inducing a plasma current. These two magnetic fields behave like a "magnetic cage" that confines the hot plasma. Unfortunately, the development of a commercial nuclear fusion reactor still requires ongoing experimental investigations, advanced physical modelling, and innovative engineering solutions. To address these challenges, artificial intelligence has been applied to solve various issues in nuclear fusion, including diagnostic data processing, plasma control, and AI-assisted engineering.A particularly promising AI approach are Physics-Informed Neural Networks (PINNs). PINNs uniquely integrate data-driven methodologies with foundational physical principles, enabling new capabilities such as meshless partial differential equation (PDE) solving, domain decomposition, multi-scale analysis, super-resolution, and direct modelling. While PINNs have been extensively studied, an optimized framework for their application to nuclear fusion is still in its infancy. Therefore, developing and proposing innovative implementations is essential for advancing efficient PINN-based solutions in plasma physics.This study demonstrates how combining prior information about physics with data-driven deep learning techniques can support the development of direct-modelling algorithms, i.e. a procedure that overcomes the process of iteratively hypothesizing a model and its parameters, followed by a numerical simulation and its validation with experimental data. Once supported by sufficiently comprehensive and realistic diagnostic information, this methodology would allow deriving realistic models that are automatically calibrated with experimental data and reflect the underlying physics of the investigated phenomena. This algorithmic solution would not only enhance performance and reliability but would also reduce the need for conventional iterative modelling techniques, contributing significantly to the advancement of nuclear fusion research.