The Application of the Novel Kolmogorov–Arnold Networks for Predicting the Fundamental Period of RC Infilled Frame Structures

The fundamental period is a crucial parameter in structural dynamics that informs the design, assessment, and monitoring of structures to ensure the safety and stability of buildings during earthquakes. Numerous machine-learning and deep-learning approaches have been proposed to predict the fundamental period of infill-reinforced concrete frame structures. However, challenges remain, including insufficient prediction accuracy and excessive computational resource demands. This study aims to provide a new paradigm for accurately and efficiently predicting fundamental periods, namely, Kolmogorov–Arnold networks (KANs) and their variants, especially radial basis function KANs (RBF-KANs). KANs are formulated based on the Kolmogorov–Arnold representation theorem, positioning them as a promising alternative to multilayer perceptron. In this research, we compare the performance of KANs against fully connected neural networks (FCNNs) in the context of fundamental period prediction. The mutual information method was employed for the analysis of dependencies between features in the FP4026 data set. Nine predictive models, including KANs, F-KANs, FCNN-2, FCNN-11, CatBoost, Support Vector Machine, and others, were constructed and compared, with hyperparameters determined by Optuna, which will highlight the optimal model amongst the F-KANs models. Numerical results manifest that the highest performance is yielded by the KANs with R² = 0.9948, which offers an explicit form of the formula. Lastly, we further dive into the explainability and interpretability of the KANs, revealing that the number of stories and the opening percentage features have a significant effect on the fundamental period prediction results.