Abstract Summary
The fast and practical computation of dynamic stiffness of foundations in layered media is a very important task to evaluate the dynamic response of critical structures. In particular, the design of wind turbines in offshore regions are very sensitive to the stiffness and damping characteristics of the foundation system. It is critical in fatigue life prediction, and ultimate limit states. In this communication, a fast assessment method is proposed via a meta-model based on Deep Learning techniques. The model is designed to detect enriched physical domain, based on Transfer Learning from low-resolution physical domain to most refined physical schemes (complex layered systems and poroelasticity). To generate the required database of dynamic stifnesses, a three-dimensional Green function for multilayered elastic and poroelastic half space in the frequency domain has been developed and adapted for source and receiver locations at the top free surface. The fundamental solution is built by potential displacements, angular Fourier transform, and radial Hankel transform. The Julia programming language has been explored to implement and optimice the computation of the Inverse Hankel Transform. A Boundary Element Method code has been developed with traction singular shape fuction for rectangular foundations. In order to consider the applicability of the Deep Learning metamodel to complex physical domains (many layers, poroelasticity), increasing the number of features, a first deep neural network has been built and trained, suitable for two-layers elastic half-space configurations. Network hyperparameters have been optimized based on error analysis. A Transfer Learing approach is explored to build deep neural networks for improved physics, including layers and poroelastic properties. Numerical tests confirm that the proposed methodology is suitable for fast computation of the dynamic stiffness of foundations, with low computing times compatible with industry requirements.
