A machine learning-based strategy to determine inductively coupled plasma robustness from non-analyte signals

The inductively coupled plasma (ICP) is a key energy source for efficient atomization, excitation, and ionization in atomic spectrometry. Plasma robustness, often represented as the Mg II/I line intensity ratio, is a well-known and commonly used diagnostic probe of energy transfer in the ICP. Typical ICP parameters are set to maximize plasma robustness with relatively large injection tube diameters, decreased nebulizer (i.e., carrier gas) flow rates, and increased radiofrequency powers. ICP optical emission spectrometry is more resistant to matrix effects at higher values of Mg II/I. We describe a strategy featuring machine learning to determine plasma robustness from non-analyte, plasma-based species of Ar, H, and O. We collected a representative dataset ( n = 4450) of plasma robustness and non-analyte signals at varying instrumental conditions and plasma environments. Tuned tree-based ensemble and neural network models yielded accurate predictions of Mg II/I robustness with root mean squared errors as low as 0.36 across the training and testing data. The trained models may be used to predict plasma robustness, given model imperfections for predictions in relatively extreme matrices. Once trained, the models do not require any physical modification to commercial instrumentation nor the addition of diagnostic elements to the plasma. This machine learning strategy with plasma-based species can be adapted to train predictive models for additional plasma characteristics. • Machine learning models trained on signals from naturally occurring, plasma-based species predict inductively coupled plasma robustness • The developed machine learning strategy requires no modification to commercial inductively coupled plasma optical emission spectrometers • A dataset ( n = 4450) of plasma robustness and non-analyte signals from varied instrumental conditions and plasma environments was used for modeling