I Congreso de Postgrado fcfm: ingeniería, ciencias e innovación

113 Santiago, 10 al 12 de agosto, 2022 SHEAR WALL BUILDING DESIG BASED ON DEEP NEURAL NETWORKS Pablo N. Pizarro 1 , 2 *, L. M. Massone 2 , F. R. Rojas 2, R. O. Ruiz2 1 Departamento de Ciencias de la Computació , Universidad de Chile, Santiago, Chile. 2 Departamento de Ingeniería Civil, Universidad de Chile, Santiago, Chile. *Email: pablo.pizarro@ing.uchile.cl ABSTRACT In the structural designof shearwall buildings, the initial process requires the interactionbetween thearchitecture and engineering teams to define the appropriate distribution of walls, a stage typically carried out through a trial- and-error procedure, without any consideration of previous similar projects. For the engineering analysis, first, the wall thickness and length definition, their location, and in some cases, the presence of new wall sections, are required to fulfill not only architectural requirements but also engineering needs such as building deformation limits, base shear, among others. The present investigation consists of two parts to help the structuring of a shear wall building: first, an artificial neural network (ANN) is used for predicting the thickness and length of the wall segments, based on information obtained from previous architectural and engineering projects, and second, two convolutional neural networks (CNN) models are used to predict new shear wall not considered by architecture. The study includes surveying the architectural and engineering plans for a total of 165 buildings constructed in Chile. The generated database has the geometric and topological definition of the walls and the slabs [1]. An ANN model was trained for the regression of the thickness and length of the wall segments of each structure, making use of a feature vector that models the variation between the architectural and the engineering plans for a set of conditions such as thickness, connectivity (vertical and horizontal), area, wall density, type of structure, type of foundation soil, the distance between elements, among other engineering parameters. The regression model (ANN) obtained results in terms of R² of 0.995 and 0.994 for the predicted wall thickness and length, respectively. As a first approach of applying an artificial intelligence model to predict the thickness and length of the walls, remarkable results were achieved; however, the initially proposed purely regressive methodology does not allow the prediction of new elements not present within the architectural plan. For this reason, another two convolutional neural network (CNN) models are used in the early process of the conceptual design of the building wall layout, fed only by architectural data (images and numerical features). The first CNN application is a regressive model that predicts the wall engineering values of the thickness, the length, the wall translation in both axes from the architectural plan, and the f loor bounding box geometrical properties such as width or the aspect ratio. The second application is a model that generates a likely image of the final engineering f loor plan to propose new structural elements not present in architecture while reinforcing the existing wall layout. In this contribution, the convolutional layers provide a better extraction of geometrical and topological features. Both models allow building a tool to predict the complete engineering f loor plan based on previously validated projects [2]. ACKNOWLEDGMENTS Beca ANID magíster Nº 22200500 y FONDECYT Regular 2020 N° 1200023. REFERENCES [1] P. Pizarro y L. Massone, Eng. Struct. 241 , 112377 (2021) [2] P. Pizarro, L. Massone, F. Rojas y R. Ruiz, Eng. Struct. 239 , 112311 (2021) I NG E N I E R Í A E S T R UC T U R A L 09