Abstract
Building a robust calibration curve is essential for accurate quantification of multicomponent mixtures. Matrix effects can distort the proportional relationship between the analyte concentration and instrumental response. In addition, classical machine learning models do not inherently incorporate simple physical constraints, such as the requirement that a zero response must correspond to a zero concentration, which can result in non-zero predictions. To address this limitations, prior-induced calibration models were proposed that inherently embed this physical constraint into the model architecture. A dataset was generated for ethanol electrooxidation products using headspace gas-chromatography-mass spectrometry (HS-GC-MS). Multiple linear regression (MLR), polynomial regression and artificial neural network (ANN) models were trained to investigate the effects of model complexity and the incorporation of physical information on predictive performance. Model selection and complexity were assessed using Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC). The prior-induced polynomial regression model achieved improved predictive accuracy while maintaining low model complexity, whereas the ANN provided comparable accuracy but was strongly penalized due to its substantially higher complexity.