Seismic Liquefaction Potential Assessed by Neural Networks

TLDR

Neural networks, which mimic brain architectures, are employed to model complex relationships between seismic and soil parameters. A simple back‑propagation neural‑network trained on field records used eight input variables—SPT, fines content, D50, τav/σ′0, σ0, σ′0, earthquake magnitude, and maximum horizontal acceleration—to predict liquefaction potential. The neural‑network model proved feasible, with performance improving as more inputs were added; the eight‑variable model—especially SPT and fines content—was most accurate and outperformed the conventional dynamic stress method.

Abstract

The feasibility of using neural networks to model the complex relationship between the seismic and soil parameters, and the liquefaction potential has been investigated. Neural‐networks are information‐processing systems whose architectures essentially mimic the biological system of the brain. A simple back‐propagation neural‐network algorithm was used. The neural networks were trained using actual field records. The performance of the neural‐network models improved as more input variables are provided. The model consisting of eight input variables was the most successful. These variables are: the standard penetration test (SPT) value, the fines content, the mean grain size D50, the equivalent dynamic shear stress τav/σ′0, the total stress σ0, the effective stress σ′0, the earthquake magnitude M, and the maximum horizontal acceleration at ground surface. The most important input parameters have been identified as the SPT and fines content of the soil. Comparisons indicate that the neural‐network model is more reliable than the conventional dynamic stress method.