A new simple neural-network based approach to predict the seismic response of levees and small height earth dams

International audienceThe issue of the seismic stability of embankments has been raised by several recent events. Given their very large cumulated length, together with the large variability of their owners, it is rarely possible to perform detailed investigations and complex numerical simulations, especially in moderate seismicity areas. Analyzing their capacity to withstand seismic loading thus requires the use of simple tools, based on a few easily available parameters. This work aims at providing an easy-to-use tool to assess peak acceleration at crest of embankments and peak acceleration of potential sliding blocks then allowing to evaluate their stability in case of earthquake. It is based on a 2D numerical parametric study combining 135 realistic configurations of embankments and natural soil layers, and four loading levels, characterized by peak acceleration at outcropping bedrock (0.01g, 0.1g, 0.3g and 0.5g). For each configuration and loading level, nonlinearity is taken into account by using equivalent shear modulus and damping consistent with the strain level (derived from a set of 1D linear equivalent computations). Artificial neural networks are then used to identify the parameters that best control the seismic response of the embankment while being easily available. They also provide a relation between these input parameters and the researched outputs (peak accelerations at embankment crest and for potential sliding blocks). A few abacus based on those neural networks are also provided as a visual tool to help engineers understanding the main trends