UNDERSTANDING THE MECHANISMS OF LIQUEFACTION-INDUCED SOIL EJECTA USING INTERPRETABLE MACHINE LEARNING

Abstract

Soil ejecta, commonly observed as sand boils, is a prominent manifestation of soil liquefaction during earthquakes, often leading to ground settlement, structural damage, and post-earthquake site instability. Despite their impact, traditional liquefaction models largely focus on triggering mechanisms and rarely address the conditions that govern ejecta formation. This study applies Rough Set Machine Learning (RSML), an interpretable rule-based algorithm, to analyze 96 historical case histories of liquefaction, including 85 cases with observed soil ejecta. Six geotechnical and seismic parameters—moment magnitude, peak ground acceleration, groundwater table depth, average critical layer depth, corrected SPT N-value, and fines content—were used as model inputs. IF–THEN rules were induced and evaluated using support, strength, certainty, and coverage metrics. Sensitivity and interaction analyses revealed the critical influence of corrected penetration resistance, groundwater conditions, and fines content on ejecta occurrence. Notably, soil ejecta was found to occur even in moderately dense or fine-grained soils, challenging conventional assumptions. The parameter interaction map and scenario map further highlighted the combined effects of seismic and subsurface conditions. This interpretable model provides valuable insights into post-liquefaction behavior and supports improved hazard assessment, mitigation planning, and resilience strategies. The findings contribute to the advancement of data-driven approaches in earthquake geotechnical engineering.