This research has aimed to investigate the possibility of applying a neural network algorithm into the structural design process of bascule bridge leaves, by creating a workflow in Grasshopper. The demand for this tool, originates from the fact that the current design process is experienced as linear and slow, and does not fit the dynamic design environment within Amsterdam.

The foundation for the generative design workflow, is a parametric model for an orthotropic bridge deck made out of steel. The model is made out of main beams, which are modeled as tubular profiles, cross beams and rib elements, which are modeled as line elements, and a bridge deck plate on top, to which all elements are welded. The parametric model is controlled by a total of 24 parameters which describe its dimensions, profiles and spacing of elements. Based on the user input for three parameters, namely the bridge deck length, bridge deck width and distance to the rotational axis, the neural network algorithm does a suggestion for the remaining parameters, based on information in the bridge database. The bridge database consists of 35 bascule bridges. A design for a steel bridge leaf is then generated. In the next step, the generated structure is forwarded into SCIA Engineer via a.xml file, for structural analysis. Within SCIA Engineer, the structural performance is assessed based on occurring Von Mises stresses under defined load combinations.

In the validation stage of the research, three steps were undertaken. In the first validation step, the predicting behavior of the neural network was optimized based on the mean squared error. The predicting behavior of the workflow was assessed using 5-fold cross validation. It appeared that the neural network had a prediction accuracy of 25.91%. Therefore, in the second validation step, the neural network’s complexity was reduced. In the improved model, the neural network predicts fourteen parameters. The total predicting accuracy of the improved model is equal to 61.07%.In the last validation step, five random cases were generated, of which their SCIA model output was compared to simplified models and the predicting behavior of the neural network was assessed. In two out of five cases, the neural network immediately suggests a good structure, while in two others, only one parameter had to be altered to create a viable structure. From validation of the SCIA models, it appeared that the moment, shear force and stress distribution for both the main beam and crossbeam showed consistent behavior through all five models. For the main beams, a simple MatrixFrame model was constructed for comparison. There were significant differences in the magnitude of occurring moments and shear forces between the MatrixFrame model and SCIA model. The magnitudes of moments, shear forces and stresses in the cross beams were compared to a hand calculation of a clamped beam. It was concluded that the SCIA models generated in the validation stage of the research, behaved correctly.

In the results section, the output of the workflow was compared to two reference cases, the Berlagebrug and Elizabeth Admiraalbrug. For the Berlagebrug, the SCIA model had such a significant error that no useful results could be extracted from this case study. For the Elizabeth Admiraalbrug, the workflow generated a structure which was approximately twice as large in steel mass. This has resulted in lower unity checks compared to the original design. A qualitative comparison between the current design process and when the created workflow would be implemented, was also done.

Based on the quantitative results obtained in the research, it was concluded that the neural network algorithm in the application developed in its current form, will not significantly improve the structural design process, due to a lack of consistency in generated results.