Comparison of Machine Learning Algorithms for Estimating the Natural Frequencies and Dynamic Response of a Planar Cantilever Truss Under Impulsive Loading

Optimal design of steel trusses subject to natural-frequency constraints requires repeated dynamic analysis over a very large number of candidate geometries and member cross-sections, which is computationally prohibitive when relying on finite element analysis (FEA) alone. In this study, the performance of three machine learning algorithms - a feed-forward artificial neural network (ANN), a random forest, and a gradient boosting ensemble - is compared for building a surrogate model that simultaneously predicts the first three natural frequencies and the peak dynamic displacement of a ten-member, two-bay planar cantilever truss under a half-sine impulsive load. A complete finite element model, comprising assembly of the truss stiffness and mass matrices, solution of the generalized eigenvalue problem for the natural frequencies and mode shapes, and modal superposition for the dynamic response, was developed for this purpose. A dataset of ۱۵۰۰ samples was generated by randomly varying the bay length, truss height, and the cross-sectional areas of three member groups (chords, verticals, and diagonals), and all three machine learning models were trained and evaluated on this dataset. The results show that the artificial neural network achieves the highest accuracy for all four output variables (coefficient of determination R² above ۰.۹۹), followed closely by gradient boosting, while the random forest yields a somewhat lower accuracy. A feature-importance analysis further reveals that the bay length is the most influential design parameter affecting both the natural frequencies and the dynamic response. These findings provide a basis for replacing repeated direct dynamic analyses in frequency-constrained truss optimization with a fast and accurate machine learning surrogate.

کلیدواژه ها:

machine learning ، planar cantilever truss ، structural dynamic analysis ، natural frequency ، artificial neural network ، random forest ، gradient boosting ، surrogate model ، modal analysis ، impulsive load