Behavior of CFRP strengthened RC multicell box girders under torsion concrete torsion;multicell box girder;carbon fiber reinforced polymer;strengthened girders;finite elements;nonlinear analysis; The use of fiber reinforced polymer (FRP) for torsional strengthening of reinforced concrete (RC) single cell box beams has been analyzed considerably by researchers worldwide. However, little attention has been paid to torsional strengthening of multicell box girders in terms of both experimental and numerical research. This paper reports the experimental work in an overall investigation for torsional strengthening of multicell box section RC girders with externally-bonded Carbon Fiber Reinforced Polymer CFRP strips. Numerical work was carried out using non-linear finite element modeling (FEM). Good agreement in terms of torque-twist behavior, steel and CFRP reinforcement responses, and crack patterns was achieved. The unique failure modes of all the specimens were modeled correctly as well.

Accurate and reliable prediction of Perfobond Rib Shear Strength Connector (PRSC) is considered as a major issue in the structural engineering sector. Besides, selecting the most significant variables that have a major influence on PRSC in every important step for attaining economic and more accurate predictive models, this study investigates the capacity of deep learning neural network (DLNN) for shear strength prediction of PRSC. The proposed DLNN model is validated against support vector regression (SVR), artificial neural network (ANN), and M5 tree model. In the second scenario, a comparable AI model hybridized with genetic algorithm (GA) as a robust bioinspired optimization approach for optimizing the related predictors for the PRSC is proposed. Hybridizing AI models with GA as a selector tool is an attempt to acquire the best accuracy of predictions with the fewest possible related parameters. In accordance with quantitative analysis, it can be observed that the GA‐DLNN models required only 7 input parameters and yielded the best prediction accuracy with highest correlation coefficient (R = 0.96) and lowest value root mean square error (RMSE = 0.03936 KN). However, the other comparable models such as GA‐M5Tree, GA‐ANN, and GA‐SVR required 10 input parameters to obtain a relatively acceptable level of accuracy. Employing GA as a feature parameter selection technique improves the precision of almost all hybrid models by optimally removing redundant variables which decrease the efficiency of the model.