Novel techniques for enhancing the performance of support vector regression chemo-metric in quantitative analysis of LIBS spectra
Laser induced breakdown spectroscopy (LIBS) is an atomic emission spectroscopy through which elemental compositions of materials can be determined with little or no sample preparation. Small sample requirement as well as it capacity for rapid and real time analysis contributes significantly to the wider applicability of the technique. However, quantitative analysis of LIBS spectra remains a challenge and requires non-linear modeling technique that fully captures the complex interactions in the laser induced plasma and ultimately reduces the effect of self-absorption. Support vector regression (SVR) recently attracts significant attention in chemo-metrics due to its sound mathematical background and unique ability to model non-linear systems with reasonable degree of precision. This work proposes two novel techniques by which the performance of SVR can be improved for the quantitative analysis of LIBS spectra. The first technique, referred to as homogeneously hybridized support vector regression (HSVR), combines two SVR algorithms in which the output of the first algorithm serves as the input to the second algorithm while the second technique, referred to as internal reference preprocessing method (IRP), uses the spectra feature that is normalized with the emission line intensity which is not significantly affected by self-absorption. The hyper-parameters of the developed models are optimized using gravitational search algorithm (GSA). On the basis of root mean square error, GSA-HSVR-WIRP (without IRP) performs better than GSA-SVR-WIRP with over 75% performance improvement while GSA-HSVR-IRP performs better than GSA-SVR-IRP with over 95% performance improvement. The outcome of this work would be very useful for precise LIBS quantitative analysis and would eventually promote wide applicability of the technique.