Cluster analysis (CA) is a multivariate tool used to organize a set of multivariate data (observations, objects) into groups called clusters. The cluster analysis method was carried out on characteristics of Ricotta cheese powder, with the effect of milk/whay ratio (formulation) and foam mat drying temperature. In this study, 4 types of formulations and 6 drying temperature were used to study the density, hygroscopic and color factors to find the formulation and optimal temperature that created the proper physical properties. The results of analysis of variance showed high temperature due to higher vapor velocity, decreased density and increased hygroscopicity (p<0.05). Also, with increasing temperature, the index "L" decreased and the indices "a" and "b" decreased. According to the results of cluster analysis, cluster 2 was selected as the best cluster for the least disparity between treatments and also due to the lowest Within-group variance. In this cluster, cheeses with a high percentage of whey in the formulation combination and low temperatures are found to foam mat drying. According to the results, the Lightness (L) of the powders of this cluster is higher, and at lower temperatures the density and hygroscopy are lower. Based on the results in general, the use of cluster analysis to select formulations for foam mat drying of ricotta cheese is a suitable method.

Artificial neural networks are a set of nonlinear equations that have the ability to adapt to establish complex nonlinear relationships between input and output variables. Artificial neural network modeling was used to predict the production of Ricotta cheese powder with the desired quality. In this study, a 4-class artificial neural network with a multilayer perceptron model was used to predict foam and Ricotta cheese powder data prepared by foam mat drying. This modeling was performed by pattern recognition method and using machine learning algorithm. Pattern recognition is the ability to recognize the order of properties or data that gives information about a system or data set. The model used for this study had 10 neurons in the hidden layer. 4 different ratios of milk and whey (treatments) were considered as input and foam density, powder density, hygroscopy, water activity, water absorption and oil absorption as model outputs. In this model, 70% of the data were used for training, 15% for testing and 15% of the data for validation. The best validation performance occurred in the 20th period. The final results showed that the model used was able to accurately predict the data related to each class with 94.8% accuracy.