Machine learning models are widely used for decades in various engineering applications, such as structural health monitoring, optimization of the properties of engineering systems or structures. For instance, in structural engineering, researchers have investigated machine learning techniques for the prediction of the natural frequencies, damage detection, and de-sign optimization of beams, frames, plates, and many other structures. Using machine learn-ing is advantageous since machine learning can reduce the cost and time consumption to solve real-world problems. These techniques do not require powerful computers and soft-ware, unlike numerical analysis methods to solve such problems. To benefit such positive as-pects of the machine learning techniques, the prediction of the first ten natural frequencies of aluminum and steel very thin, thin, and thick beam structures under fixed-free, fixed-sim-ply supported, and simply supported boundary conditions by using Radial Basis Function Regressor, Random Forest Regressor, Multilayer Perceptrons Regressor, and Support Vector Machine Regressor with Pearson VII Universal Function Kernel (Puk) has been presented. The dataset required for the analysis is obtained via the Finite Element Analysis considering Euler-Bernoulli and Timoshenko Beam Theories. The performance of the machine learning models has been investigated and compared by examining (i) the thickness-length ratio, (ii) boundary conditions, and (iii) natural frequencies of the beam structures. Results indicate that the considered regression machine learning models are effective in predicting the natural frequencies of beam structures. Among all four regression machine learning models, Support Vector Machine Regressor with Puk and Random Forest models are robust and accurately predict the natural frequency values of the structures by an average accuracy of 98.78% and 98.88% regardless of the boundary conditions and thickness-length ratio of beam structures. On the other hand, Radial Basis Function Regressor and Multilayer Perceptron Regressors predict the first ten natural frequencies by 96.36% and 94.17%, respectively.