ANN based LIBS models for quasi-experimental spectra relevant for materials for next-step fusion reactors

Abstract

Following the successful demonstration of machine learning (ML) models for laser induced breakdown spectroscopy (LIBS) adaptation in fusion reactor fuel retention monitoring using synthetic data [Gąsior et al., Spectrochim. Acta, Part B 199, 106576 (2023)], this study focuses on implementing operability on experimental data. To achieve this, Simulated Eperimental Spectra (SES) data are generated and used for validation of a chemical composition estimation model trained on dimensionally reduced synthetic spectral data (DRSSD). Principal component analysis is employed for dimensionality reduction of both SES and DRSSD. To simulate real experimental conditions, the synthetic data, generated by a dedicated tool [M. Kastek (2022), “SimulatedLIBS,” Zenodo. http://dx.doi.org/10.5281/zenodo.7369805] is processed through the transmission function of a real spectroscopy setup at IPPLM. Separate and optimized artificial neural network models are implemented for conversion and chemical composition estimation. The conversion model takes DR-SES as features and DR-SSD as targets. Validation using converted SES data demonstrates chemical composition predictions comparable to those from synthetic data, with the highest relative uncertainty increase below 40% and a normalized root-mean-square error of prediction below 7%. This work represents a significant step toward adapting ML-based LIBS for fuel and impurity retention monitoring in the walls of next-generation fusion devices.