Enhancing concrete compressive strength prediction using Kolmogorov–Arnold networks

Abstract

In construction projects, reliably forecasting the compressive strength of concrete samples plays a critical role. To enhance the prediction accuracy, this research introduces a prediction model for concrete compressive strength utilizing Kolmogorov–Arnold Networks (KANs). Based on two datasets with different scales, containing 1030 and 324 groups of concrete samples, respectively, the KANs model was used for modeling, and the optimal hyperparameters of the KANs model were identified using grid search. The results show that, on the two datasets, the KANs model achieved a root mean square error of 5.09 MPa and 0.97 MPa, and the goodness of fit reached 0.90 and 0.99, respectively, outperforming the comparison methods multi-layer perceptron’s neural networks and support vector machine. Under simulated measurement noise conditions (Gaussian perturbations with 10% feature-level variance), the KANs model demonstrated robust performance preservation, exhibiting <5% relative degradation in prediction accuracy. This resilience to instrumentation-level uncertainties substantiates the model’s operational reliability in practical concrete testing environments. To assess the generalization ability of the KANs model, the model was trained with a small proportion of samples from dataset 2 based on dataset 1. The results indicate that when the sample proportion was increased to 30%, the KANs model achieved an RMSE and R² of 2.85 MPa and 0.9, respectively, in the remaining 70% of the samples in dataset 2, demonstrating the model’s good generalization performance.