A data-driven fusion convolutional neural network framework for predicting frequency-dependent microwave heating performance across foods

The complementary-frequency shifting strategy has been developed as a promising approach to improve heating uniformity in solid-state microwave processing. However, an extensive frequency-sweeping stage is needed to collect all frequency-dependent thermal patterns, consuming approximately 20 % of the total heating time (roughly 1 min of a 6-min heating cycle). This study introduces a data-driven fusion convolutional neural network (CNN) model that can predict multiple frequency-dependent thermal patterns requiring only one pattern collected as a seed, effectively eliminating the time-consuming frequency-sweeping stage in the dynamic frequency shifting process. The optimized model showed good prediction accuracy with SSIM around 0.9 and RMSE values of 0.86 °C. The trained model is versatile and can be seamlessly adopted across various seed frequencies, food materials, and geometries. Statistical analysis indicates that the model performance is not affected by the seed frequency used or food materials involved, while the food geometries not represented in the training dataset may result in reduced prediction accuracy. Nevertheless, the model can still properly locate the hot and cold spots regardless of those potential affecting factors, which is crucial for determining complementary patterns for shifting. Demonstrated by Multiphysics simulations, the optimized model applied to dynamic frequency shifting shows more stable temperature change during heating, potentially enhancing food quality preservation.