Artificial intelligence and machine learning can help organize commercial dog food data and generate candidate wet food recipes, but gaps in ingredient and nutrient information limit its use for complete formulations, according to research results published in the journal Informatics.
Researchers in Kazakhstan developed a system combining text analysis, ingredient recommendations and mathematical optimization. Scientists at Astana IT University and S. Seifullin Kazakh Agrotechnical Research University conducted a computational study to evaluate product classification and similarities among generated recipes.
The system drew on 330 commercial dog foods from Hill’s Pet Nutrition, breed and disease databases, and nutrient information for 126 ingredients selected from a larger food composition dataset. Both dry and wet commercial products informed the analysis, while recipe generation focused on wet foods.
Users entered characteristics including breed, body weight, age, activity level and reproductive status. The system calculated energy requirements using the European Pet Food Industry Federation’s 2024 nutritional guidelines and selected ingredients associated with a chosen dietary purpose. A mathematical solver then calculated ingredient proportions using limits on ingredient categories and ranges for moisture, protein, fat and carbohydrates observed in commercial wet foods. These calculations did not establish complete nutritional adequacy.
The researchers noted that individual characteristics primarily affected caloric requirements and feeding quantities. The selected dietary specialization determined the recipe’s structure, so changing a dog’s weight or activity level did not necessarily change its ingredient combination.
A major task was filling gaps in product descriptions. Of the 330 products, 242 initially lacked functional category tags, such as kidney care or weight management. After manual annotation, the researchers evaluated two automated labeling approaches. Both methods struggled with overlapping or broadly worded categories, including aging care and joint support.
The researchers also compared two approaches to ingredient selection. Machine learning generated recipes with more overlapping ingredient sets, while statistical association methods produced greater variety. These comparisons measured recipe consistency and diversity rather than nutritional quality.
Several limitations remain consequential for pet food formulation. Commercial ingredient proportions were unavailable, and generic descriptions required substitutions. For example, the researchers mapped “meat and animal by-products” to chicken and “cereals” to rice. These were modeling assumptions rather than verified ingredient identities.
The ingredient database also lacked specialized components such as hydrolysates and functional additives. Although the paper describes calculating micronutrient requirements, its discussion and conclusions acknowledge that vitamins, minerals and technological additives were not fully incorporated into recipe optimization.
The authors cautioned that “the simulated recipes may deviate from production standards or veterinary requirements.” However, the work demonstrates a possible foundation for organizing product information and exploring candidate formulations.
