Helium retention feature in the boron deposited layer on tungsten substrate by laser-induced breakdown spectroscopy and machine learning approach

ITER is designed for a burning plasma operation in which Tungsten (W) tiles are used as the first wall and diverter materials. Studying He dynamics in B-coated W wall is essential to understanding the effect of boronized plasma-facing components in fusion reactors, as these post-conditioning materials significantly influence helium retention and release. Plasma-wall interactions (PWI) are an important issue in ITER fusion reactors. PWI would lead to wall erosion and impurity redeposition. As a product of the D-T burning plasma, helium (He) ash would be retained or co-deposited on plasma-facing components (PFCs), affecting the stable operation of burning plasma. Laser-induced breakdown spectroscopy (LIBS) is proposed as a promising in-situ diagnostic approach for monitoring D/T and He retention and impurity deposition on PFCs of tokamak devices. In this study, the LIBS technique was used to investigate the helium retention feature in the boron (B) layer on tungsten substrate in 10⁻⁵ mbar. Five He-retention samples on the boron deposition layer on tungsten substrates were prepared by pulsed laser deposition (PLD) method at different ambient pressures in our lab. The investigations indicate that the atomic spectral line of helium (He-I 587.56 nm) was observed in the spectra of the first three laser shots. The depth profiles of He, B, and W in the boron-deposited layer on tungsten substrate were performed by LIBS to determine the co-deposition layer thickness. The concentration of He in the co-deposition layer samples measured by TDS is 7 × 10²⁰ He/m². The plasma parameters, such as plasma electron temperature and electron number density, were calculated to validate the local thermodynamic equilibrium. Machine learning (ML) algorithms are used to classify the co-deposits and substrates. The first three principal components (PC1, PC2 & PC3) of the unsupervised ML algorithm (PCA) give a classification accuracy of 91.7 %. The supervised ML algorithm neural network achieved training and testing accuracy of 100 % and 96.7 %, respectively.