Evaluation and updating of Ishihara’s (1985) model for liquefaction surface expression, with insights from machine and deep learning

Abstract

Liquefaction surface-manifestation is a popular proxy of damage potential for infrastructure. Models for predicting it are thus commonly used, and often codified, in earthquake engineering practice. One such model is that of Ishihara (1985) who proposed empirical “H1–H2” curves considering the influence of the non-liquefied crust on surface expression. Yet, while widely used and cited, these curves were trained on just ∼300 data points from two earthquakes. Accordingly, this study evaluates and updates the Ishihara (1985) model using 14,400 data points from 24 earthquakes, while also comparing against three other manifestation models from the literature. In addition to retraining the H1–H2 model via traditional regression, new variants are developed via machine- and deep-learning. Each of the new H1–H2 models outperforms the original in unbiased testing and is suitable for application. Ultimately, however, this paper also explores the limits of H1–H2 models and the apparent inefficiency and/or insufficiency of their predictor variables. In this regard, the models developed herein may perform better than any other, yet new models are still needed to account for factors influential in producing surface manifestation in a more explicit and mechanistic manner.