Use of convolutional networks in the conceptual structural design of shear wall buildings layout

In the structural design of shear wall buildings, the initial process requires the interaction between the architecture and engineering teams to define the appropriate distribution of the walls, a stage typically carried out through a trial-and-error procedure, without any consideration of previous similar projects. In previous work, a database of 165 Chilean residential projects of reinforced shear wall concrete buildings was built, which fed a regressive neural network model to predict the wall's engineering thickness and length values from an architectural 30-feature input vector, which accounts for geometric and topological properties, archiving remarkable results regarding the coefficient of determination (R2). However, a regressive model of this nature does not incorporate a spatial detail or contextual information of each wall's perimeter, and also, the prediction of other parameters such as the wall translation has a poor performance. For this reason, the present research proposes a framework based on convolutional neural network (CNN) models to generate the final engineering floor plan by combining two independent floor plan predictions, considering the architectural data as input. The first plan prediction is assembled using two regressive models that predict the wall engineering values of the thickness, the length, the wall translation on both axes from the architectural plan, and the floor bounding box width and aspect ratio. The second plan prediction is assembled using a model that generates a likely image of each wall's engineering floor plan. Both independently predicted plans are combined to lead the final engineering floor plan, which allows predicting the wall's rectangles design parameters and propose new structural elements not present in architecture, making the methodology an excellent candidate to accelerate the building wall layout's early conceptual design.