Machine learning-based prediction of statically equivalent seismic forces in pin-supported cylindrical reticulated shells

Reticulated shells exhibit complex vibrations during earthquakes, encompassing components in horizontal and vertical directions, and multiple vibration modes occur. In particular, single-layer reticulated shells with a small depth relative to their span exhibit many vibration modes, and the shapes of these modes can vary depending on the geometry. The method for rapidly setting equivalent static seismic forces remains unexplored. In response to the above background, this study proposes a novel approach for calculating the seismic forces on single-layer reticulated shells using machine learning techniques. The shells in focus are pin-supported cylindrical reticulated shells, typically for the roofs of gymnasiums used as evacuation facilities during severe earthquakes in Japan. Machine learning uses numerical analysis results for approximately 20,000 shells, with varied spans, half-open angles, and aspect ratios. A method for preprocessing the principal vibration modes as image data is proposed, after which the imaged vibration modes are predicted from the shape parameters of the shell using a neural network. The prediction accuracy is analyzed, and a method for rapidly calculating seismic loads based on combining predicted vibration modes is proposed. These seismic loads are compared with the response spectrum method results, and their effectiveness is discussed.