Explainable neural network models for rapid prediction of dielectric constant and acid value in edible oils using...

Reliable evaluation of edible oil quality is essential for ensuring consumer safety and detecting market adulteration. This study presents a rapid and non-destructive approach that integrates Fourier Transform Infrared (FTIR) spectroscopy with chemometric and deep learning techniques to predict key quality parameters of edible oils. The dielectric constant (DC) and acid value (AV) were selected as target indicators to assess oil degradation. Principal component analysis (PCA) effectively distinguished pure and market oil samples through well-defined clustering patterns, reflecting underlying compositional differences. Conventional chemometric models, including partial least squares regression (PLSR), competitive adaptive reweighted sampling-based PLSR (CARS-PLSR), and random forest-assisted PLSR (RF-PLSR), demonstrated satisfactory predictive performance, with coefficient of determination (RHossain et al. (2025)) values exceeding 0.80 for both DC and AV. To further enhance prediction accuracy and capture nonlinear relationships within spectral data, advanced deep learning models such as one-dimensional convolutional neural networks (1D-CNN), residual networks (ResNet), and CNN-LSTM architectures were developed. Among these, the 1D-CNN model achieved superior performance, yielding R [2] values above 0.95 along with lower RMSE. To improve model interpretability, explainable artificial intelligence techniques were employed. SHAP (SHapley Additive exPlanations) and Gradient-weighted Class Activation Mapping (Grad-CAM) provided complementary insights into the spectral features driving model predictions. These analyses identified chemically meaningful FTIR regions associated with key functional groups, including ester carbonyl (C=O), aliphatic CH, unsaturated CC, CO stretching, and OH vibrations linked to oxidation products. Overall, the proposed framework offers a robust, rapid, and reliable strategy for edible oil quality assessment and adulteration detection.