The phenomenon of soil liquefaction in saturated, cohesionless deposits remains one of the most critical seismic geotechnical hazards, profoundly impacting the structural resilience of civil infrastructure. Under cyclic earthquake loading, the rapid generation of excess pore water pressure under undrained conditions leads to a precipitous decline in effective stress (σ′) and the subsequent degradation of soil shear strength. This loss of bearing capacity and stiffness poses severe risks to underground transit systems, buried lifelines, and near-surface foundations. Consequently, accurate assessment of liquefaction potential is an indispensable component of performance-based seismic design. Recently, machine learning algorithms, particularly Artificial Neural Networks (ANNs), have emerged as robust computational tools capable of mapping the highly nonlinear, multidimensional constitutive behavior of soils, offering enhanced predictive capabilities for seismic hazard microzonation.
This study presents a comprehensive evaluation of liquefaction susceptibility along the alignment of Line 2 of the Tabriz Metro, advancing a predictive framework driven by an optimized ANN model. The classical stress-based assessment framework evaluates liquefaction hazard by correlating the seismic demand, defined as the earthquake-induced Cyclic Stress Ratio (CSR), with the soil’s inherent capacity to resist triggering, termed the Cyclic Resistance Ratio (CRR). To establish a rigorous analytical baseline, stratigraphic, hydrogeological, and in situ testing data from 54 boreholes were synthesized. The regional seismic demand was characterized by generating a Peak Ground Acceleration (PGA) zonation map, derived through a finite-fault stochastic simulation approach. Utilizing these site-specific PGA parameters, the spatial liquefaction potential was initially computed via two widely adopted deterministic frameworks: the Standard Penetration Test (SPT)-based simplified method by Idriss and Boulanger (2010), using corrected blow counts (N1)60, and the geophysics-based methodology by Andrus et al. (2004), utilizing small-strain shear-wave velocity (Vs​) as the primary proxy for cyclic resistance.
The analytical outputs of these semi-empirical procedures were subsequently leveraged as reference datasets to train and validate the AI model. A feed-forward “fitnet” ANN architecture was programmed in MATLAB, employing a nonlinear sigmoid activation function within the hidden layer and a linear transfer function at the output node. The geotechnical database comprised 249 distinct records. To ensure robust model generalization and prevent algorithmic overfitting, 149 records were allocated to the network’s development phase (partitioned into 89 for training, 30 for validation, and 30 for internal testing). Crucially, a hold-out set of 100 independent records was reserved exclusively for blind prediction. Preliminary parametric sensitivity analyses optimized the network topology, establishing an architecture with 10 hidden neurons to balance computational complexity with predictive fidelity.
The results demonstrate that the optimized ANN framework yields high-accuracy predictions that strongly converge with the conventional Vs​-based deterministic assessments. The neural network successfully mapped the complex spatial variability of critical Vs ​thresholds, implicitly capturing the influence of geotechnical nuances such as soil aging and diagenetic cementation on liquefaction resistance. These findings indicate that ANNs can function as highly efficient, reliable surrogate models for macro-scale liquefaction hazard assessment. However, an inherent epistemic limitation of the current model is its reliance on algorithmically generated pseudo-labels derived from conventional methodologies rather than empirical post-earthquake observational data. To engineer universally resilient predictive frameworks, future investigations must prioritize training algorithms on comprehensive, global databases of documented liquefaction case histories, thereby minimizing epistemic uncertainty across diverse seismotectonic environments.