Enhancing prediction stability and performance in LIBS analysis using custom CNN architectures.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Laser-Induced Breakdown Spectroscopy (LIBS) is a powerful technique for elemental fingerprinting, offering rapid, simultaneous multi-element analysis.
• Existing LIBS analysis often overlooks plasma temperature and electron density variations, impacting accuracy.
• Bridging the gap between simulated and real data analysis is crucial for robust LIBS applications.

Purpose of the Study:

• To develop and compare predictive models (PLS and CNNs) for elemental concentration in LIBS spectra.
• To account for variations in plasma temperature and electron density in LIBS data analysis.
• To enhance the accuracy and stability of elemental concentration prediction using advanced machine learning.

Main Methods:

• Utilized Partial Least Squares (PLS) and Convolutional Neural Networks (CNNs) for predictive modeling.
• Trained and tested models on simulated LIBS data, evaluating performance using Root Mean Square Error of Prediction (RMSEP).
• Applied pre-trained models to real LIBS spectra and fine-tuned CNN architecture for element-specific prediction.

Main Results:

• CNNs achieved superior performance with median RMSEP values below 0.01, outperforming PLS (0.01-0.05) on simulated data.
• Real LIBS spectra analysis consistently identified Aluminum (Al), Silicon (Si), and Iron (Fe) with high predicted concentrations.
• Modified CNNs highlighted the impact of regularization, sample weighting, and custom loss functions, revealing patterns for Ca, Mg, Zn, Ti, and Ga.

Conclusions:

• CNNs offer enhanced stability and predictive accuracy for elemental concentration in LIBS analysis compared to PLS.
• The study successfully bridged the gap between simulated and real LIBS data analysis by accounting for key plasma parameters.
• Fine-tuning CNNs provides a pathway to prioritize and detect specific elements within complex LIBS spectra.