LA-TReQNet: Improving Multielement Quantification Model for Laser Ablation Inductively Coupled Plasma Mass Spectrometry Based on Deep Learning Network

Laser ablation inductively coupled plasma mass spectrometry
(LA-ICP–MS)
is a widely used quantitative technique. However, conventional calibration
approaches that depend on internal standards, external reference materials,
and linear regression are subject to practical limitations. To address
these challenges, we present LA-TReQNet, an end-to-end deep learning
framework, that enables fully automated quantitative calibration in
LA-ICP–MS. A CNN-LSTM architecture was trained on a data set
of 221,364 labeled mass spectra from 5676 samples. We demonstrate
that preprocessing of mass spectrometry data substantially impacts
model performance and propose an optimized strategy incorporating
power transformer-based standardization and data set grouping. By
training on extensive multielement data sets, the deep learning model
captures complex empirical relationships, thus establishing for the
first time a standard-free calibration approach. The optimized LA-TReQNet
model accurately quantified 39 elements in three RMs (BCR-2G, BHVO-2G,
and BIR-1G) from independent laboratories, confirming its robustness
to data source variations. When applied to a broader set of RMs, LA-TReQNet
achieved deviations of 0.2% ± 5.8% (SD, n =
198) for major elements and −0.9% ± 9.2% (SD, n = 898) for trace elements relative to certified reference
values. The results confirm that deep learning-based elemental quantification
achieves accuracy on par with conventional approaches while eliminating
the reliance on internal or external standards, thus significantly
expanding the applicability of LA-ICP–MS to complex and diverse
samples. Moreover, LA-TReQNet eliminates data quality fluctuations
from researcher experience differences in traditional handling, improving
both processing efficiency and result stability.