Speaker
Description
Integrated data analysis is essential for the full exploitation of diagnostic measurements. Physics-Informed Neural Networks (PINNs) present a compelling alternative to standard methods (e.g. the Bayesian statistics). As a novel branch of artificial intelligence, PINNs seamlessly integrate data-driven methodologies with physical equations, offering a highly efficient approach.
PINNs provide distinct advantages over traditional methods, including the ability of handling incomplete physics equations, managing noisy data, and operating independently of mesh constraints. This work focuses on evaluating the potential of a Physics-Informed Neural Network (PINN) algorithms to reconstruct plasma equilibrium using a multi-diagnostic approach incorporating magnetics, kinetic pressure, and interferometer-polarimeter data.
In tokamaks, equilibrium reconstruction is inherently an ill-posed problem. To achieve accurate results, it is essential to constrain algorithms with multiple diagnostic inputs. Consequently the PINN technology allows deriving a more accurate and complete overview of the plasma state in all phases of the discharges. Additionally, the redundancy provided by the multi-diagnostic approach enables simultaneous equilibrium reconstruction and "diagnostic of diagnostics" by readily identifying outliers or faulty measurements and effectively filtering out noise.
Results are presented for both synthetic data [1], [2] (using TokaLab, a virtual tokamak developed at the University of Rome "Tor Vergata" for education and research) and data from JET including experiments from the recent DT campaigns [3].
References
[1] N. Rutigliano et al., “Physics-informed neural networks for the modelling of interferometer-polarimetry in tokamak multi-diagnostic equilibrium reconstructions,” Plasma Phys. Control. Fusion, vol. 67, no. 6, p. 065029, Jun. 2025, doi: 10.1088/1361-6587/addde6.
[2] N. Rutigliano, A. Murari, P. Gaudio, M. Gelfusa, and R. Rossi, “Optimisation of Physics-Informed Neural Network Architecture and Training for Tokamak Equilibrium Reconstruction,” Plasma Phys. Control. Fusion, Mar. 2026, doi: 10.1088/1361-6587/ae54c9.
[3] N. Rutigliano, A. Murari, P. Gaudio, M. Gelfusa, and R. Rossi, “Multi-diagnostics reconstruction of magnetic equilibrium and kinetic profiles using Physics-Informed Neural Networks with applications to JET,” Nuclear Fusion, Feb. 2026, doi: 10.1088/1741-4326/ae4916.