On the applicability of neural networks for soil dynamic amplification analysis

Artificial Neural Networks (ANN) have gained a solid status as a tool for modeling complex phenomena in different areas of research and engineering practice. In this paper, their applicability to estimate the mapping of seismic acceleration from bedrock to free surface in a complex soil profile is explored. Such a use is intended to serve as a hypothesis-free alternative to the dynamic amplification analysis, which is currently based on geophysical and soil dynamics procedures. Were the neural networks to be useful to such a mapping, they could in principle be employed for several purposes, such as soil identification using instrumental data, design of early warning systems and estimation of probabilistic spectra via Monte Carlo simulation, in which the ANN act as an efficient solver surrogate. The conditions under which these ambitious purposes can be reached are discussed. Two classes of multi-layer perceptrons were tested, which are characterized by time-independent and time-dependent connections. It is shown that the first class of networks is useful for response spectrum mapping, while the second performs very well in the assessment of free-surface time series. It arises as the main conclusion that the most promising perspective of application of ANN in this respect is for the estimation of probabilistic free-surface spectra, which is an important goal for the modern trend of reliability-based aseismic design. The limitations to the other said applications are also highlighted.