GNN-LSTM-based fusion model for structural dynamic responses prediction

With the rapid growth of deep learning technology, the potential for its use in structural engineering has substantially increased in recent years. This study proposes an innovative deep-learning fusion network architecture based on the graph neural network (GNN) and long short-term memory (LSTM) network. The proposed fusion model can accurately predict nonlinear floor acceleration, velocity, and displacement responses of typical steel moment-resisting frames (SMRF) with 4 through 7 stories subjected to strong ground motions. The fusion model framework overcomes a major drawback in existing deep learning models as it can predict the dynamic responses of various structures. This has widened the range of potential applications of structural surrogate models generated through deep learning technology on the design and analysis of building structures with different geometries. Additionally, this paper presents two LSTM-optimized learning strategies, namely packing padded sequences and sequences compression strategies. These strategies improve the model's performance significantly without modifying its architecture. Finally, this study reveals that the model's internal graph embedding is highly correlated with certain critical structural parameters, such as the first natural period and the height of a building. This shows that the proposed fusion model is interpretable and has the ability to extract vital information that influences structural dynamic responses.