Integrating graph neural networks with physics-informed loss function for mechanical response prediction of hollow concrete structures with morphed honeycomb configurations
Hanmo Wang, P. Tan, Yee Zin Foo, Shawn Owyong, Zhong-Rong Lu, Huan Ting Chen, Tam H. Nguyen, Alexander Lin
Abstract
• Efficient assessment of mechanical behavior in hollow concrete components. • A simulation simplification approximates 3D model into 2D model. • A Graph Neural Network (GNN) to process and analyze complex structures. • A customized loss function comprising of physics-informed loss terms. In modern construction, hollow concrete structures with customized configurations, such as morphed honeycombs, offer an impressive high stiffness-to-weight ratio, maintaining structural rigidity while minimizing material usage. Yet, evaluating these structures requires laborious simulations, like Finite Element Methods (FEMs), which impedes efficiency. In this research, we proposed a hybrid model that integrates a Graph Neural Network (GNN) and a physics-informed loss function for swift and precise structural analysis of hollow concrete structures with morphed honeycomb configurations. The training data for the proposed model and the benchmark model was generated through computational simulations calibrated by mechanical test results for concrete honeycomb specimens with porosities of 33%, 45%, and 55%, which were cast using silicon molds created from 3D-printed polymer formwork. The hybrid model in the present study outperforms the benchmark GNN by enhancing accuracy: it reduces the Mean Square Error (MSE) of nodal displacement from 5.37E to 5 to 1.91E-5 and decreases the MRE of structural stiffness from 0.0478 to 0.0273. This hybrid model can be integrated into the workflow of structural optimization, serving as an efficient evaluation tool. Initially focusing on structural analysis, the hybrid model also holds potential for other applications such as thermal analysis.