The article addresses a pressing issue in applying machine learning methods (specifically, artificial neural networks) to structural calculations: the shortage of experimental data for training data-driven models. An approach is proposed for forming synthetic training data by means of numerical modeling with the finite element method in the ANSYS software package. As a demonstration, a forward problem of determining the required area of longitudinal reinforcement in a simply supported reinforced concrete beam is presented. The features of modeling the behavior of materials (concrete, steel, and composite reinforcement) are discussed, and the assumptions made and the choice of failure criteria are substantiated. An algorithm for data generation is described, including the parameters and their variation ranges, as well as a procedure for filtering the results according to code-based strength and serviceability criteria. A large synthetic dataset has been generated, comprising about 15,000 simulation cases, which can be used for training and testing machine learning models. It is concluded that employing numerical simulation to obtain reliable and representative training datasets can significantly reduce the cost and duration of physical experiments.