Formulation of special fats by neural networks: A statistical approach

In the present work, a neural network able to formulate Cats with three ingredients derived from soybean tone refined oil and two hydrogenated base stocks) was built and trained. The training of the network was accomplished with data on the solid fat content (SFC) of 112 products, association with the proportions of the raw material used in their formulation. After the training, the network furnished, from the requested solid profiles, the possible formulations for the desired product. According to the statistical analysis applied to the results obtained, larger mean errors were observed in products with very low SFC and the smallest errors were found in products with high SFC. Regarding different temperatures, the network performance was more accurate for 10, 20, and 25 degrees C than for 30, 35, and 37.5 degrees C, where the lower measurements resulted in larger relative errors. According to evaluation by industrial experts, all the responses furnished by the network after its training were considered within the acceptable variation limits. For these experts, the network knowledge generalization (accomplished with products not presented during the training) was considered highly efficient (nearly 100%).