This paper presents the optimization of concrete mixtures composition related to a physical property and the process of production of trial mix design by using the multi-layered feed-forward neural networks. This optimization was conducted because there is no clear method of designing concrete mixture composition and for the purpose of shortening procedure of the trial mix design of concrete. Mix design depend on many variables and deterministic models cannot give good results. The goal of the research was to make a model of a neural network, on the set of available data from 288 trial mix, which would, with highest accuracy, predict the compressive strength of concrete at the age of 28 days. In order to attain as high accuracy of obtained results as possible, three levels of input data to the neural networks were considered. On each of the applied groups of input data, the neural networks with 1 and 2 hidden layers were formed. On the basis of the adopted neural network, an algorithm for usage of the network in actual situations was made, applied on an actual model.