Deep learning approach to structural performance prediction of tall buildings in seismic areas

To study the interplay between architectural form and seismic response in tall buildings with outer diagrids during the early design phase, advanced computational analysis and artificial intelligence methods are adopted. Dynamic numerical simulations of moderate magnitude earthquakes are performed, investigating different architectural forms to inform early design decisions within an extensive design space. Parametric design software, e.g. Rhinoceros, Grasshopper, is considered together with OpenSees. The numerical analysis evaluates the response to vertical static loads and lateral seismic excitations, in terms of inter-storey drifts, displacements and various stress metrics, under seven ground motion scenarios, using a dataset of 1000 tall buildings selected through Latin hypercube sampling. A multi-input multioutput feed-forward neural network is developed for this purpose; fine-tuning parameters such as network architecture, learning rate, batch size, and validation techniques enable the model to achieve a low validation loss, indicating an effective learning process. Designers can take advantage of the workflow applied to the database to guide their early decisions.