Benchmarking in Laser-Induced Breakdown Spectroscopy

abstrakt

angličtina

The recent technological boom in LIBS resulted in the production of very large spectroscopic data.1 Various processing techniques and methods have been developed over time with ranging applicability and performance. Well-established algorithms based on classical statistics are not anymore usable for more advanced processing of large high-dimensional data. On the other side, modern Machine Learning techniques (Neural Networks, Support Vector Machines, etc.) are very often overused or applied in an incorrect way. Establishing a robust benchmark for a specific task (classification or quantification,...) is necessary to distinguish between approaches and select a “correct” solution/s to each problem. We are presenting a challenging benchmark for material classification through LIBS spectra. It consists of 138 physical samples, separated into 12 categories according to their elemental composition. For each sample 500 spectra of dimension 40,002 wavelength values are available (in training part of the dataset). Later, extended version of the benchmark (5000 spectra per sample) will be released.

Laser-Induced Breakdown Spectroscopy; LIBS; machine learning; benchmarking;

POŘÍZKA, P.; VRÁBEL, J.; KÉPEŠ, E.; KAISER, J.

18. 1. 2020

Tucson, Arizona

241

242

2

http://icpinformation.org/Winter_Conference.html