Background: Accurate prediction of punching shear capacity is critical for the safety and reliability of reinforced concrete (RC) flat slab structures, yet conventional empirical code provisions often exhibit significant scatter. While Machine Learning (ML) models, such as Artificial Neural Networks (ANN), have shown promise, their performance is highly dependent on effective hyper-parameter tuning, a process often neglected in prior studies.

Objective: This study aims to develop a novel, highly robust, and accurate predictive model for the punching shear capacity of RC slabs by integrating an ANN with a systematic Metaheuristic Optimization approach.

Methods: An extensive experimental database was compiled from the literature. A Genetic Algorithm (GA) was employed to optimize the key hyper-parameters (e.g., network architecture, learning rate) of the ANN model, creating a GA-ANN hybrid model. The GA's fitness function was defined to minimize the Mean Absolute Error (MAE) on a dedicated validation set. Model performance was evaluated on an independent testing set using statistical metrics, including, MAE, and RMSE, and compared against non-optimized baseline models and established design codes.

Results: The GA-ANN model achieved significantly superior predictive accuracy on the testing set ( of 0.957 , MAE of 14.5 kN) compared to the baseline ANN and conventional code methods. The optimization process successfully determined a globally efficient set of hyper-parameters, resulting in notably reduced scatter and bias in the prediction-to-actual ratio. Comparative analysis demonstrated the model’s CoV (0.110) was substantially lower than ACI 318 (0.295) and Eurocode 2 (0.225), proving its uniform reliability across various material and geometric ranges.

Conclusion: The integration of metaheuristic optimization, specifically the Genetic Algorithm, provides a powerful and necessary framework for developing highly reliable machine learning models in structural engineering. The resulting GA-ANN model offers a superior, data-driven alternative for the robust and efficient estimation of punching shear capacity in RC slabs.