Machine Learning-Based Prediction of ICP-MS Results Using XRF Data

HYEJIN SHIN, Taeyeon Kim, Seung-Muk Yi, Kwon Ho Jeon, Kyungmi Lee, Hyeog Ki Chae, Seoul National University, Seoul, Korea

     Abstract Number: 321
     Working Group: Source Apportionment

Abstract
X-ray fluorescence (XRF) and inductively coupled plasma mass spectrometry (ICP-MS) are widely used techniques for trace element analysis. While XRF offers rapid, non-destructive analysis with minimal sample preparation, ICP-MS provides higher sensitivity and lower detection limits. Due to their complementary characteristics, the two methods are selectively applied depending on analytical objectives and sample matrices. This study investigates the feasibility of predicting ICP-MS results using XRF data through machine learning techniques. By identifying patterns between XRF and ICP-MS data, predictive models to estimate ICP-MS values from XRF measurements could be developed.

In situations where ICP-MS measurements are limited or unreliable—such as for arsenic (As), which may suffer from sample preparation artifacts—XRF data can serve as an alternative for indirect estimation. For example, As and lead (Pb), which are frequently co-emitted from sources like coal combustion, often exhibit strong correlations. Given that Pb is reliably detected by XRF, it can act as a surrogate predictor for As, along with other XRF-accessible elements.

To evaluate the validity of the machine learning-based reconstruction, Positive Matrix Factorization (PMF) was applied to both the original ICP-MS dataset and the reconstructed dataset generated from XRF-based predictions. Source profiles and contributions derived from each dataset were compared to determine the consistency and reliability of the reconstructed values.

Additionally, previous studies using dithiothreitol (DTT) assays to assess oxidative potential of particulate matter reported inconsistent sets of DTT-correlated elements depending on the analytical method (XRF vs. ICP-MS). In this study, we further examine the correlation between DTT activity and the machine learning-predicted ICP-MS values to assess whether the reconstructed data can recover physiochemically relevant associations with oxidative toxicity.

Acknowledgement

This study was supported by the National Institute of Environment Research, funded by the Ministry of Environment (NIER).

This research was supported by Particulate Matter Management Specialized Graduate Program through the Korea Environmental Industry & Technology Institute (KEITI) funded by the Ministry of Environment (MOE).