A GENERIC METHOD FOR DETERMINING MOISTURE SORPTION ISOTHERMS OF CEREAL GRAINS AND LEGUMES USING ARTIFICIAL NEURAL NETWORKS

Moisture sorption isotherms (MSI) are required to optimize handling, drying, processing and storage of food products. MSI are obtained by solving nonlinear equations iteratively, which is normally a difficult process. A generic approach was successfully used for collective prediction of MSI for 12 cereals and five legumes simultaneously, by taking advantage of the superior computational capabilities of artificial neural networks (ANNs). A total of 779 observations were collected from literature and used in training and validation of ANNs. The ANNs model used product type (grains or legumes), sorption state (adsorption or desorption), temperature and equilibrium moisture content as inputs to predict equilibrium relative humidity. A one- and two-hidden-layer ANNs were implemented. The overall prediction results obtained from ANNs were found to compare well with those reported for conventional analytical MSI models. The average prediction results obtained from the two-hidden-layer ANN for mean square error, deviation modulus and R2 were 0.009, 4.47% and 0.984, respectively. The results for each product compared well with those obtained from analytical MSI models. Four important issues that are commonly raised when implementing ANNs in prediction of MSIs were explained. These included preventing network over-fitting, minimizing the time and effort needed for ANN architecture optimization, validating reproducibility of the results and validating the network capability to predict new observations. Unlike conventional MSI models, ANNs used the same network to predict MSIs for several products simultaneously. The principle can be easily expanded to predict MSI for other larger sets of various products, which can save time and effort. Practical Applications Determination of moisture sorption isotherms (MSI) is required for optimization of grain drying, processing, handling and storage operations. They can be also used to evaluate theoretical drying energy requirements and optimum storage conditions for a specific food product. In addition, they are used in food engineering calculations related to equipment design, shelf life evaluation and stability during storage operations. Determination of MSI for food products involves the use of conventional nonlinear MSI models, which requires an iterative solution methodology. The results are also specific to the food product investigated. Artificial neural networks were therefore proposed in this study as an alternative method for the collective prediction of MSI for some cereal grains and legumes. The method proposed in the study was accurate and reliable. In addition, it can be extended to larger varieties of food materials.