Salima Ait El Hocine

Earthquake-induced soil liquefaction represents a severe geohazard causing catastrophic infrastructure failure in prone coastal zones, requiring an advanced environmental spatial assessment for their sustainable land-use planning. This study utilizes advanced computational intelligence models to predict earthquake-induced soil liquefaction in Boumerdès, Algeria, an area heavily affected by the 2003 (M_w 6.8) earthquake. Utilizing a comprehensive subsurface database of 1984 geotechnical records encompassing lithology, hydrogeological configurations, and seismic parameters, advanced deep learning frameworks are developed and optimized via automated Neural Architecture Search (NAS). The continuous Factor of Safety (F_s) is calculated to distinguish stable profiles from vulnerable strata, benchmarking conventional ANN and DNN models against NAS-optimized variants (NAS-ANN and NAS-DNN) using a stratified 5-fold cross-validation scheme. The optimized hybrid NAS-DNN framework effectively captured non-linear soil responses, achieving a training correlation coefficient ($R_{train}$) of 0.9518, a validation coefficient ($R_{validation}$) of 0.8843, and a cross-validated mean $R$ of approximately 0.82, demonstrating improved predictive reliability compared to traditional models. Ultimately, this optimal network is embedded into the ‘GeoLiquefy-AI (v1.0)’ interface. To ensure reliability for safety-critical applications, we integrated a SHAP explainable AI framework, validating the model’s geomechanical logic by mapping physical soil-liquefaction dependencies. This deployment-ready tool enables rapid, transparent hazard calculations, providing a scalable platform for seismic microzonation and proactive urban risk mitigation.