Canadian Journal of Chemical Engineering General Chemical Engineering
Physics‐informed surrogate modelling for intermittent wheat drying: Fractional kinetics, Arrhenius generalization, and cross‐condition prediction
Abstract
Abstract Intermittent grain drying induces non‐Markovian, sub‐diffusive moisture transport that classical Fickian and curve‐by‐curve empirical models cannot intrinsically generalize. This work establishes a unified k ( T , I ) structural requirement, k ( T , I ) = k 0 exp.(− E ₐ/RT)· I ᵝ, that simultaneously governs empirical thin‐layer models, the fractional Mittag‐Leffler parametric model, and a fractional physics‐informed machine learning (fPIML) framework. Experiments on BRS‐Jacana wheat at three temperatures (40, 55, and 70°C) and two intermittency periods (5 and 10 min) yielded 152 gravimetric observations on eight drying curves. Two‐stage globalization of six empirical benchmarks confirmed Arrhenius activation energies of 22–42 kJ mol −1 ( R 2 ≥ 0.91) for the physically well‐behaved models (Page, Logarithmic, Midilli), while the fractional‐parametric model yielded E ₐ = 33.9 kJ mol −1 and a global law α ( T , I ) ≈ −1.28 + 803.4/Tᵅ − 0.246·ln(I) ( R 2 = 0.844). The fPIML embeds the Caputo fractional governing equation Dₜᵅ(MR) + λ ᵅ( T , I )·MR = 0 into the training loss, with λ ᵅ structured as a learned Arrhenius expression. Under thermal leave‐one‐temperature‐out validation, the fPIML (RMSE = 0.0479, R 2 = 0.955) achieves lower RMSE and higher explained variance than both the integer‐order PINN (RMSE = 0.0604, R 2 = 0.929) and the data‐only ANN (RMSE = 0.0795, R 2 = 0.875). Under the critical intermittency extrapolation test, trained exclusively on I = 5 min and tested on I = 10 min, the fPIML achieves the lowest MAE (0.059) among all models; the integer‐order PINN reaches R 2 = 0.504 but with an MAE 90% higher (0.112), while the data‐only ANN collapses to R 2 = −0.025.