Interpretable Machine Learning in Food Analysis: Closed-Form Chemometric Calibration of Mid- and Near-Infrared Spectra Using Kolmogorov–Arnold Networks

A Kolmogorov–Arnold network (KAN) learns a univariate function on every edge, so a trained network can be written as a closed-form equation, unlike a multilayer perceptron (MLP). A KAN was compared with five models, with selection nested within each withheld-lot or withheld-season fold, on tahini adulteration by attenuated total reflectance Fourier-transform infrared (ATR–FTIR) spectroscopy (1554 spectra, released here) and mango dry matter by near-infrared (NIR) spectroscopy (across-season test). No accuracy advantage was found: the KAN’s coefficient of determination (R2) was comparable to a parameter-matched MLP’s but never higher on average, and on withheld lots no difference between the leading models was significant. Most of the accuracy lost in converting the network to an equation was due to the conversion procedure; correcting it raised the median external R2 of the tahini equation from 0.952 to 0.988. Repeated conversions of one network gave equations scoring −0.76 to 0.96, whereas an elementary edge basis makes the printed equation identical to the network up to rounding. With every lot serving as the blank, detection limits for samples resembling the calibration population were 8.9–21.6% adulterant, and blends at 4–12% were not quantified. The KAN provides an exact, inspectable screening calibration rather than higher accuracy.