Investigating the Potential of Data-Drivenseismic Response Analysis: Integrating Neural Networks with Dynamic Mode Decomposition

This study introduces a new approach to seismic response analysis that integrates Dynamic Mode Decomposition (DMD) and Neural Networks (NN) takes advantage of the strengths of each method. NNautonomously capturing complex nonlinear phenomena, but their lack of interpretability limits their use in geotechnical engineering, where transparency and reliability are critical. DMD, on the other hand, provides interpretability through the use of linear operators, but has limited adaptability when dealing with nonlinear dynamics. By combining these methods, we aim to create an interpretable data-driven model that balances adaptability and reliability. The proposed hybrid approach effectively decomposes the seismic response characteristics into a linear term (captured by the DMD) and an input-dependent nonlinear term (captured by the NN). We will examine whether the model learns the unrepresentable relationships of the DMD in one-dimensional seismic response, and whether DMD-based analysis is feasible. This integrated framework will not only advance the use of NN in engineering applications, but also establish the basis for reliable and interpretable ground response analysis. By compensating for the limitations of each method, the hybrid model allows for efficient seismic response modelling and more accurate inverse analysis of subsurface properties based on surface-based observations. The findings of this study elucidate the fundamental mechanisms underlying the high performance of NNs and highlight the essential considerations for developing data-driven models that can be applied effectively and confidently in an engineering context.

Keywords: Neural networks, Dynamic mode decomposition, Machine learning, Seismic response analysis.