Data-Driven Structural Modeling
During the planning phase of slender structures such as skyscrapers, wind turbines, and bridges, optimizing their dynamic structural properties using computer-based models is essential. To reduce the high computational effort required by these models, surrogate models are employed that capture the behavior relevant to the analysis of structures or their components with sufficient accuracy while excluding irrelevant details. For instance, artificial neural networks offer promising data-driven approaches for this purpose. However, these techniques generally require a large quantity and high quality of data, which poses a complex challenge, particularly in civil engineering. Therefore, we are researching innovative solution methods, such as active learning techniques, which use optimization algorithms to collect data adaptively and more efficiently.
(a) Accuracy of the data-driven model of a nonlinear single-degree-of-freedom oscillator after training with data obtained under various experimental conditions. The developed active learning method identifies the optimal experimental conditions for the model using a minimal number of data samples. (b) Model result after data collection using the Sobol method. (c) A better model result is achieved with the developed method.
Milicevic P, Altay O (2024). Data generation framework for inverse modeling of nonlinear systems in structural dynamics applications. Acta Mechanica. 235(3):1493-1515. https://doi.org/10.1007/s00707-023-03532-3
DFG-Project: Data-Driven Nonlinear Modelling and Control of Structures