Machine learning‐based wind‐induced response analysis in rectangular building models with limbs

Abstract

SummaryThis study investigates the impact of different positions of two limbs on the structural response of a rectangular building model to wind forces. The building's geometry assumes Z and + shapes based on specific limb configurations. Computational fluid dynamics (CFD) simulations are performed to quantify wind‐induced pressures, resulting in wind force coefficients. These coefficients are used to develop predictive machine learning models through Gene Expression Programming, Group Method of Data Handling‐combinatorial (GMDH‐Combi), Model Tree, and Artificial Neural Network (ANN) techniques. The ANN analysis explores various combinations of training algorithms, adaptation functions, activation functions, and performance functions to enhance model accuracy. Among these, the Levenberg–Marquardt (LM) with gradient descent with momentum (GDM) adaptation function and sigmoid activation function exhibit superior performance with high R‐squared values. These predictive models are then employed for a comprehensive comparative assessment of the maximum wind force coefficient (CF, max) concerning different limb positions and angles of attack (AOA). For CF, max vs Limb position, variations of limb position are examined for most critical cases of AOA. Similarly, the study of CF, max vs AOA involves an exhaustive investigation into the variation of AOA for the building with the worst limb position. The analysis reveals that changes in AOA have a more pronounced effect on CF, max compared to alterations in limb position. Interestingly, within the AOA range of 1.5 to 2.5, the CF, max consistently reaches a minimum across all models. However, the relationship between CF, max and the critical structural parameter ‘S' (representing limb position) remains less conclusive for the most significant AOAs.