Microwaves have been well recognized as a high-efficiency and clean energy source, and are being widely used for various types of food processing applications, including reheating (Chen et al., 2014), tempering (Li et al., 2018), defrosting (Pitchai et al., 2016), drying (Duan et al., 2010; Wray & Ramaswamy, 2015), pasteurization and sterilization (Chandrasekaran et al., 2013), and so on. However, owing to the inherent dielectric heating concept of microwaves, using microwaves results in fast but nonuniform heating performance with hot and cold spots formed in the heated foods. This nonuniform heating can cause issues such as extreme moisture loss at overheated regions, texture degradation, and compromised food quality and safety, leading to failure to achieve consistent product standards and concerns among customers regarding the reliability of microwave-powered processes (Michalak et al., 2020; Pitchai et al., 2014; Su et al., 2022).

A major reason for such a nonuniform heating phenomenon by microwave is the standing wave patterns correlated to the use of constant microwave frequency by the magnetron that generates a fixed distribution of electromagnetic fields, and thereby, unchanged thermal patterns (Datta & Anantheswaran, 2001; Luan et al., 2017). The solid-state technique, which allows the alteration of frequency during heating, can serve as a promising replacement to the current magnetron for emitting microwave energy with changing frequency (Yang et al., 2022a; Zhou, Pedrow, et al., 2023). By varying the microwave frequency, different thermal patterns of hot and cold spots can be generated to improve the overall heating uniformity (Du et al., 2019; Yang et al., 2022a). Various frequency shifting strategies have been reported to improve the heating uniformity, where utilizing the complementary patterns by different frequencies would yield significantly better heating uniformity than shifting the frequencies in order (Yang, Fathy, Morgan and Chen, 2022a, Yang, Fathy, Morgan and Chen, 2022b; Yang & Chen, 2024).

To apply the complementary frequency shifting strategy to a realistic microwave processing scenario, it is necessary to obtain the frequency-dependent thermal contributions (patterns) for determining the complementarity among frequencies. Currently, these frequency-dependent patterns have been collected either before a real microwave heating process (Yang et al., 2022a; Zhou, Tang, et al., 2023) for concept demonstration purposes or through a time-consuming frequency sweeping stage during a real-time microwave heating process (Yang et al., 2022b). For the collection of frequency-dependent thermal patterns before real microwave heating, this approach was mainly used for concept demonstration to show microwave frequency affects heating patterns, which was not applicable in practice since real food products have great variations of material properties, shapes, and mass. While for the collection of frequency-dependent patterns during microwave heating using extensive frequency sweeping, the sweeping process takes a significant amount of time of the whole heating process, for example, 1 out of 4–6 min microwave heating was used for sweeping as reported by Yang et al. (2022b). This online approach, even slightly compromises the pattern accuracy as it does not account the dielectric change along the heating, has been proven to be effective as it can be easily adopted to heating scenarios with various foods, accounting the slight variation among replicates (Yang et al., 2022b). However, in a typical domestic microwave process, which heating typically lasts 4 to 6 min, sweeping all frequencies to obtain patterns would require at least 1 min based on the frequency sweeping rates demonstrated in solid-state microwave processes (Yang et al., 2022b; Yang et al., 2023). Hence, to better implement the complementary-frequency shifting method, reducing the time required for pattern collection is necessary, which would allow more time for strategically shifting the complementary frequencies.

A convolutional neural network (CNN) model is a powerful tool that can be used to conduct image processing in agriculture research (Ciocca et al., 2018; Liu et al., 2021; Saptono et al., 2024; Yurdakul et al., 2024), mainly for classification or object detection tasks. In addition to the commonly used CNN architecture, which includes convolutional layers for feature extraction and fully connected layers for regression, the UNet model, which is composed entirely of convolutional layers, has been widely adopted for image segmentation tasks, particularly in the field of medical imaging (Baccouche et al., 2021; Dolz et al., 2018; Ronneberger et al., 2015). The widespread adoption of UNet in medical imaging is attributed to its encoder-decoder structure with skip connections that preserve spatial resolution during processing, making it particularly suitable for microwave thermal pattern prediction where accurate spatial localization of hot and cold regions is critical for determining frequency complementarity. Given the property of 2-dimensional (2D) output from such an encoder-decoder-based model, it is possible to conduct microwave pattern prediction where the output from the model is a temperature profile at the top surface of the food. In this project, we propose to employ a data-driven fusion UNet-based CNN framework, whose input consists of information from multiple types of sources, to correlate the change of thermal patterns along with frequency, realizing the prediction of patterns from one-to-all frequencies. This work demonstrates the feasibility of a data-driven fusion CNN framework. The framework is trained and tested using simulated thermal patterns generated across different frequencies for foods with varying dielectric properties and geometries. The objectives of this study are to:1)

Implement a baseline UNet-based CNN model;

2)

Optimize the CNN model hyperparameters and structure, for investigating the model performance;

3)

Evaluate and visualize the performance of the optimized model.