Pilot Study: Machine Learning and Deep Learning Study for Fluid Structure Interaction Problems
-
2022-07-29
Details:
-
Creators:
-
Corporate Creators:
-
Corporate Contributors:
-
Subject/TRT Terms:
-
Publication/ Report Number:
-
Resource Type:
-
Geographical Coverage:
-
Edition:Project draft report 08/01/2022
-
Corporate Publisher:
-
Abstract:Coastal bridges are critical to emergency response after extreme events and are vulnerable to cascading seismic-tsunami events. After the 2011 earthquake and subsequent tsunami in Japan, instances of damage and collapse were observed in Japanese bridges that survived the earthquake but failed under the hydrodynamic loads induced by the tsunami. The Pacific Northwest in the United States could experience similar tsunami hazards. To ensure reliable mobility after extreme events, it is necessary to understand, model, and design bridge response for tsunami loading. However, studies on wave-structure interaction are constrained by the financial cost of experiments and the computational cost of computational fluid dynamics (CFD) and fluid-structure interaction (FSI) simulations. To practically perform such simulations with reduced computational cost, a pilot study that used machine learning algorithms for basic structural engineering problems is presented. Similar machine learning models can eventually be used to estimate the tsunami loading on bridges based on structural properties and flow conditions. Machine learning (ML) and deep learning (DL) algorithms, when trained for a specific problem, can produce faster results than finite element methods (FEM). Nonetheless, ML and DL algorithms are data-driven and could produce unreliable results when evaluated outside the training data domain. The interpretability of ML and DL algorithms can also be lost during the training of the model. The reliability and interpretability of ML and DL can be resolved by introducing physics into the ML and DL architectures. This project studied the performance of data-driven and physics-informed DL algorithms in structural engineering applications. The DL algorithm was studied for static and dynamic problems using single-degree-of-freedom (SDOF) oscillators representing a simplified model of a bridge pier. Physics was introduced into the DL algorithm by extracting the residual from the finite-element analysis framework, OpenSees, and integrating it with the loss function during training. The performance of the DL algorithm with and without physics was evaluated by using different loss functions, activation functions, and other hyperparameters. For an SDOF for linear static and linear dynamic problems, the data-driven and physics-informed deep learning algorithms produced similar results. Moreover, if an appropriate neural network architecture was utilized, the DL models were able to extrapolate well beyond the test data. Although the studied cases were for relatively simple SDOF linear static and dynamic problems, DL algorithms have the potential to produce reliable results for multi-degree-of-freedom systems, including the relevant physics. The approach of introducing OpenSees along with the ML and DL algorithms also presents an opportunity for engineers to produce fast and reliable results by supplementing analyses with ML and DL techniques. Nonlinear problems, multiple-degrees-of-freedom systems, and FSI studies, including the residual during the learning process, should be assessed to evaluate the performance of DL algorithms beyond the simple structural systems presented herein.
-
Format:
-
Funding:
-
Collection(s):
-
Main Document Checksum:
-
Download URL:
-
File Type: