ARTIFICIAL INTELLIGENCE AS A TOOL FOR DEVELOPING RECIPES FOR PECTIN-CONTAINING PASTE

Abstract

The article explores the possibility of using artificial intelligence (AI) technologies to optimize and develop pectin paste recipes. The topic's relevance is due to the growing demand for functional food products with improved technological characteristics and the need to accelerate the development of innovative recipes.

The work aims to create and test a method for using machine learning algorithms to predict pectin pastes' rheological, organoleptic, and physicochemical parameters and optimize their recipe composition. Mathematical modeling methods, neural networks, genetic algorithms, regression analysis methods, and statistical data processing were used to achieve this goal. A database of 147 experimental recipes was formed, which includes complete data on the component composition and relevant quality indicators of the finished product. The study used apple and citrus pectin with different degrees of esterification, sweeteners, acidity regulators, and structuring agents. A complex model based on artificial neural networks was developed, considering the influence of 14 independent variables (recipe components) on eight output product quality parameters. The applied multilayer architecture of the neural network with two hidden layers provided a coefficient of determination (R²) at the level of 0.943 for the rheological properties of the product.

The high determination index indicates the constructed model's reliability and significant predictive ability. Additionally, the model was validated through cross-validation and the deferred sampling method. The proposed system automatically determines the optimal ratios of ingredients and generates new recipe combinations with predicted properties. The created software complex has a modular structure and an intuitive interface, which allows food technologists without special knowledge in AI to use the system effectively.

 Experimental verification showed high prediction accuracy: the average error between the predicted and experimental values (R²) of structural and mechanical indicators was 4.7%, organoleptic – 6.2%, and physicochemical – 3.9%. A synergistic effect of combining different types of pectin was revealed, which reduces the total concentration of functional ingredients while maintaining the specified product properties. The developed system also allows for predicting the shelf life of finished products depending on the selected recipe and storage conditions. Using the developed system reduced the development time of new recipes by 73% and reduced the number of necessary laboratory experiments by 68%.

The practical value of the work lies in creating innovative tools for the food industry, which allows for the quick adaptation of the recipes of pectin-containing pastes to the specific needs of production and consumers.