Unsupervised Inference of Hydrogen Isotope Ratios from Balmer-α Spectroscopy In Tokamak Plasmas Via Physics-Informed Neural Networks (PINN)

摘要: Accurate determination of the hydrogen isotope ratio is essential for particle inventory control and tritium accounting in fusion devices. Conventional spectral fitting approaches based on iterative least-squares optimization are computationally intensive and prone to convergence failure under poor initial guesses, while purely data-driven supervised networks require extensive labeled datasets and exhibit limited extrapolation to experimental conditions outside the training distribution. We report an unsupervised Physics-Informed Neural Network (PINN) that embeds a differentiable 18-component Gaussian forward model directly into the inference architecture. The model accounts for Zeeman splitting, Doppler broadening, instrumental broadening, and three distinct neutral populations. The framework retrieves eight physical parameters, including X H (hydrogen-to-deuterium ratio, X H =n H /(n H +n D )), from raw Balmer-α spectra without reliance on ground-truth labels. A composite loss function enforces multi-view reconstruction consistency and spectral shape regularization, while applying supervised constraints on the magnetic field and neutral temperatures. Training on 360,000 synthetic spectra converges within 40 epochs, yielding a validation loss of 0.03496 and a maximum spectral residual of 0.006. Inversion on 180,000 independent test samples achieves R 2 =0.9999 for X H prediction. Under 1%-3% synthetic photon noise, integration with a Kalman filter doubles the inversion precision; Allan deviation analysis indicates a detection limit of 3.07-7.5 × 10 -5 at 10 s integration time. The proposed approach will offer a label-free, millisecond-scale inference capability compatible with real-time isotope ratio monitoring on current and next-step tokamak devices, with its reliability and accuracy having been successfully validated on experimental data from the EAST tokamak.