A study on quantitative analysis methods for mixed solution concentrations based on deep learning | JoVE Visualize

Area of Science:

Analytical Chemistry
Spectroscopy
Machine Learning

Background:

Accurate ion concentration determination in mixed salt solutions is crucial for various industrial and environmental applications.
Traditional methods often struggle with spectral overlap in complex mixtures, limiting non-destructive analysis.
Near-Infrared Spectroscopy (NIRS) offers a promising non-destructive approach, but requires advanced data analysis for complex samples.

Purpose of the Study:

To develop and validate an efficient, non-destructive deep learning model for predicting ion concentrations in multi-component salt solutions using NIRS.
To compare the performance of the proposed model against traditional chemometric methods (PLSR, RF).
To evaluate the impact of spectral preprocessing (water subtraction vs. raw spectra) on model accuracy.

Main Methods:

A Convolutional Neural Network (CNN) integrated with an attention mechanism was developed for NIRS data analysis.
Preprocessing included outlier removal, smoothing differentiation, and feature selection.
The CNN-Attention model was trained and tested on NIRS data from three-component mixed salt solutions (NaCl, KCl, MgCl2).
Model performance was compared with Partial Least Squares Regression (PLSR) and Random Forest (RF) using raw and water-subtracted spectra.
SHapley Additive exPlanations (SHAP) was used for model interpretability.

Main Results:

The CNN-Attention model significantly outperformed PLSR and RF models, particularly with raw spectral data.
Prediction coefficients of determination (R2) for K+, Na+, and Mg2+ were 0.939, 0.897, and 0.853, respectively.
Models built on raw spectra showed superior predictive performance compared to those using water-subtracted spectra in cases of severe spectral overlap.
SHAP analysis confirmed the model's rationality and the effectiveness of selected features.

Conclusions:

The proposed CNN-Attention model effectively addresses spectral overlap challenges in NIRS analysis of complex salt mixtures.
Deep learning, specifically the CNN-Attention architecture, demonstrates significant potential for accurate quantitative analysis of multi-component solutions.
This approach offers a robust and non-destructive method for ion concentration determination.