A Hybrid Deep Learning and Machine Learning Approach for Predicting Aqueous Solution Concentrations

Abstract

Assessing solution concentration is essential across multiple scientific disciplines; however, it is often complicated by limitations in instrument precision, sample impurities, and environmental variables. Low concentration levels frequently necessitate sophisticated methods such as spectroscopy or chromatography, which require specific apparatus and expertise. Conventional methods might be laborious and occasionally inadequate for accurate measurements. Consequently, researchers continually develop better, more efficient, and economically viable methodologies. Recent technological advancements, including deep learning and machine learning, facilitate the development of efficient, cost-effective systems for determining solution concentration levels, applicable to environmental monitoring and food safety tests. Therefore, this research developed a methodology for estimating solution concentrations through deep learning feature extraction and machine learning-based prediction. Images of the solution at varying concentrations were used to train models that apply deep learning for feature extraction. Linear regression (LR), artificial neural network (ANN), support vector regression (SVR), and random forest (RF) were then evaluated for using the extracted features to forecast the concentrations. Using features extracted from Visual Geometry Group 16-layer Convolutional Neural Network (VGG16) with LR, ANN, SVR, and RF yielded absolute prediction errors of 0.056229, 0.080000, 0.112172, and 0.026640, respectively, for concentration class prediction (classes 1–10). When the concentrations of classes 1 to 10 were evenly changed from 0 ppm to 4500 ppm, using VGG16 to extract features and RF to predict concentrations resulted in an average absolute error of 13.32 ppm, an RMSE of 0.072531 (normalized class scale) and 36.31 ppm (concentration scale), and an R² of 0.999361. The findings indicated that the proposed inexpensive method could efficiently classify the solution in different concentration classes and forecast their concentrations.