Researchers in Kazakhstan have built an artificial-intelligence and machine-learning platform that creates wet dog-food formulas tailored to a pet’s breed, weight, age and activity level. The interdisciplinary team combined expertise in computer science and animal nutrition to design a system capable of handling many variables at once.
System design and data sources
The model draws on information from 330 commercial dog foods supplied by a major pet-nutrition brand, alongside breed and disease databases and nutrient details for 126 ingredients taken from a broader food composition set. Those data streams provide the foundation for matching nutritional targets with ingredient options.
Users enter the animal’s characteristics, including reproductive status, and the platform then estimates energy needs using the European Pet Food Industry Federation’s 2024 nutritional guidelines. After the energy estimate is set, the algorithm selects ingredients that correspond to a chosen dietary purpose, such as weight control or joint health.
A mathematical solver allocates ingredient amounts while respecting category limits and the typical moisture, protein, fat and carbohydrate ranges found in wet products. The resulting calculation produces a candidate formula, but it does not verify full vitamin, mineral or additive compliance, leaving that step for later validation.
Computational assessment of product categories and ingredient selection
The study, conducted by researchers at Astana IT University and S. Seifullin Kazakh Agrotechnical Research University, analyzed 330 commercial dog foods from Hill’s Pet Nutrition together with breed and disease databases and nutrient data for 126 ingredients. Both dry and wet products contributed to the analysis, but the recipe-generation process was limited to wet foods.
Read Also: Polish pet food firm enters Slovakia
Two automated labeling approaches were applied after manual annotation of 242 products lacking functional tags, and both faced difficulty with broadly defined categories. For ingredient selection, a machine-learning method produced recipes with more overlapping ingredient sets, whereas a statistical association technique yielded greater variety; the evaluation focused on recipe consistency and diversity rather than nutritional quality.
Data gaps and labeling hurdles
One major obstacle was the absence of precise ingredient proportions in the commercial listings. The team had to replace generic terms, such as “meat and animal by-products” with chicken and “cereals” with rice, based on modeling assumptions drawn from standard industry practices.
Specialized components like hydrolysates and functional additives were also missing from the database, limiting the system’s ability to address specific health claims that rely on those ingredients.
Potential impact and next steps
Nevertheless, the authors warn that simulated formulas may diverge from production standards or veterinary requirements, emphasizing the need for complete ingredient and nutrient data before commercial use. Further work will focus on expanding the ingredient database, incorporating regulatory limits and testing the generated recipes in real-world manufacturing settings.
