CONICET | Buscador de Institutos y Recursos Humanos

Laser-induced breakdown spectroscopy (LIBS) is a promising technique that provides spatiallyresolved elemental composition data through imaging capabilities [1]. However, analysing complexLIBS datasets remains challenging due to the intricate spectral and spatial patterns present [2]. Theobjective of this study was to develop and apply deep learning approaches for advanced analysis of LIBS imaging data. A variety of samples including steels, zirconium and aluminium were analysed using a commercial LIBS system together with a web microscope camera. A total of 1000 LIBS images were acquired as shown in Figure 1, depicting the elemental emission distributions across the sample surfaces [1]. A convolutional neural network (CNN) was employed to extract the spatial features embedded within the images. The CNN was trained on a subset of labelled LIBS images to learn the relationships between image patterns and targets, such as elemental identification [2]. The trained CNN model achieved high classification performance (>95%) for differentiating steel,Fig 1: Typical raw image taking zirconium and aluminium based on their characteristic LIBSby camera spectra [2]. Additionally, the CNN showed promise for the identification and compensation of spectral interferences, indicating potential for improved quantitative analysis. Hyperparameter optimization established a three-block CNN architecture incorporating max pooling and dropout as optimal for this application.

enviar mensaje