Neural networks for the pendulum test with twin-tire impactor on flat and cylindrically curved glass panels

Abstract Flat and cylindrically curved glass panels are part of modern architecture. Such panels must be able to withstand, among other loads, accidental human impacts. The proof against soft body impact can be carried out using the pendulum test with a twin-tire impactor. Experimental proofs are time-consuming and expensive, whereas numerical proofs can only be carried out using suitable software and a validated model. The present investigations aim to provide an easy-to-use, yet accurate, method for calculating the pendulum test on such glass panels. Therefore, an extensive numerical parameter study has been conducted on concave, flat, and convex glass panels with different glass thicknesses, impact energies, and impact points. Polynomial data fitting would result in a significant loss of accuracy; therefore, neural networks are used. Based on the results of the parameter study, eight neural networks were trained for different glass thicknesses and impact energies. It is demonstrated that neural networks can reproduce the pendulum test with high accuracy, serving as a simple alternative to complex calculations. In contrast to numerical simulations, the application of neural networks is straightforward and can be done with nearly any software, as it requires only basic matrix operations. The primary results are the maximum principal stresses and maximum deformations of the glass panels, which can be used for the proof of single glass units and insulating glass units according to the standard in use. The neural networks are provided as freely accessible and easy-to-use Python code as supplemental data to this paper.