Ceramics-Silikáty 69, (3) 443 - 456 (2025)

With the increasing demand for the increasing performance of ultra-high-performance concrete (UHPC) in engineering construction, accurately predicting its compressive and tensile strength and optimising the material mix design has become a research focus. This paper proposes a hybrid model combining a multilayer perceptron (MLP) and LightGBM, which integrates the deep feature extraction capability of MLP and the efficient regression capability of LightGBM to achieve the high-precision prediction of UHPC compressive and tensile strength. Experimental data under different w/c (0.18, 0.19, 0.20, 0.22), curing temperatures (40 °C, 60 °C, 80 °C), and an ageing period of 56 days were collected for the model training and validation. The results show that the hybrid model outperforms the individual models, particularly exhibiting a high generalisation capability at low w/c, with R₂ reaching 0.98 in the validation and test sets and a mean absolute error (MAE) of only 1.02 MPa. Finally, the effects of different mix proportions and curing temperatures on the model′ s prediction results are discussed, providing valuable reference data for UHPC material design and engineering applications.