A heat-mass transfer coupled physics-informed neural network for optimizing radio frequency pasteurization uniformity of oat seeds

Abstract

Oat seeds are vulnerable to Salmonella contamination, making effective pasteurization technology essential for ensuring microbial food safety. Radio‑frequency (RF) pasteurization is promising for foodborne pathogen control, but poor heating uniformity hinders its commercialization. Traditional optimization relies on time‑consuming experiments or finite‑element simulations, accurate but computationally expensive. Embedding microbial inactivation kinetics directly into multi‑physics solvers further increases computational cost. According to reported correlations between thermal‑moisture and pasteurization uniformity for low‑moisture foods, this study established a physics‑informed neural network (PINN)‑based framework for RF pasteurization to optimize coupled electromagnetic‑thermal‑mass transfer performance. The PINN surrogate model enforced Maxwell’s equations, the heat conduction equation, and Fick’s moisture diffusion equation using multi‑physics simulation datasets. Simulations and optimizations terminated at a volume‑averaged moisture content of 8 % (w.b.), (safestorage level). Two‑stage Bayesian optimization tuned PINN hyperparameters and yielded optimal RF conditions (electrode gap: 149 mm, material height: 41 mm) at the pasteurization endpoint. The trained PINN predicted temperature, moisture, and electric‑potential fields to compute simulation‑based temperature (TUI) and moisture uniformity indices (MUI). Compared with non‑optimized groups, the experimental heating uniformity index (λ_T) was 22.14%‑42.70% lower, while the experimental moisture uniformity index (λ_M) improved by 26.81%‑49.43%. Experimental validation of RF pasteurization against Salmonella Enteritidis PT 30 showed optimized parameters reduced the pasteurization uniformity index (PUI, λ_P) by 21.7 %, indicating improved heat‑mass uniformity yields more uniform microbial inactivation. PINN automatic differentiation revealed that optimal parameters reduced coefficients of variation (CoV) in electric field, polarization, and thermal/moisture gradients. This study provides a computationally‑efficient, physically interpretable approach for RF pasteurization optimization of low‑moisture granular foods.