Deep Learning-Based Keras Network Formulation for Predicting the Shear Capacity of Squat RC Walls and Sensitivity Analysis

Squat-reinforced concrete (RC) shear walls with an aspect ratio of less than two are commonly used as lateral load-resisting buildings. It is frequently utilized in nuclear power plants and building structures due to its lateral strength and high stiffness. It is distinguished by its optimal cost and excellent performance. Nonetheless, precise assessment of the shear strength of squat shear walls is crucial for design specifications, and its computation can be exceedingly variable and intricate due to several efficient, expensive, and time-consuming constraining elements. The present study utilizes Keras deep learning techniques to develop a model for predicting the shear strength of squat RC walls to find a way to overcome these issues. The most comprehensive dataset of 1424 RC squat wall test specimens collected from the published literature to date has been used to develop the proposed deep learning model as well as three well-known machine learning models: RF, ANN, and LR. The results demonstrated that the Keras network exhibited a lower error rate and higher accuracy when predicting the shear strength of squat walls compared to earlier machine learning models, achieving 97.3% accuracy compared with the highest value in the RF algorithm, reaching 93.4%. Furthermore, parametric and sensitivity analyses were performed to verify that the algorithms can identify the most significant variables significantly influencing shear strength. The results showed that the (hw) was the most influencing factor on the peak shear strength of the squat shear wall as a ratio (6.36%), according to the results of the sensitivity analysis, followed by (lw) as a (5.10%), (tf) (4.96%), ( f´c ) (4.69%), (tw) (4.06%), (fy) of the web as a ratio (3.94%), and (ρh) (3.89%). These results and analyses were obtained using the (KNIME) analytics platform software, characterized by its vital role in precise computing operations and simple handling without the need for codes to reduce costs and time, and it was supported for Python and R languages.