Reconstruction of plasma equilibrium and separatrix using convolutional physics-informed neural operator

In this paper, we present PlaNet (PLAsma equilibrium reconstruction NET work), a Deep -Learning -based tool for performing plasma equilibrium and separatrix reconstruction considering both magnetic and non-magnetic measurements. PlaNet consists of three Neural Networks (NNs) running at three stages: (i) Equilibrium Net, the physics -informed neural operator accounting for the accurate reconstruction of the poloidal flux map; (ii) XPlim Net, which classifies if the equilibrium is diverted or limiter and computes the poloidal flux at the separatrix; and (iii) Separatrix Net, which gives the (r, z) coordinates of the separatrix starting from the outputs of the previous stages. The plasma equilibrium solver at the basis of this tool exploits a convolutional physics -informed neural operator, which solves the plasma equilibrium problem satisfying the Grad-Shafranov Equation by including a physics -informed term in the loss function. To constrain the equilibrium and achieve a better reconstruction, especially in the core region of the plasma, PlaNet uses also non-magnetic measures, such as the pressure profile or the poloidal current profile. PlaNet has been trained and validated on a dataset of 81986 synthetic equilibria of an ITER -like device, showing to be an effective tool for fast and accurate equilibrium ad separatrix reconstruction.