Machine Learning for Liquefaction Hazard Mapping: A Case Study for New Zealand

Earthquake-induced soil liquefaction poses significant risks to urban infrastructure in seismically active regions. Recent events, notably the 2011 Christchurch earthquake in New Zealand, demonstrate that liquefaction-induced damage can exceed that from ground shaking. This emphasises the need for scalable liquefaction hazard assessment tools. Traditional assessment methods that rely on cone penetration tests (CPT) and standard penetration tests are impractical for large-scale applications (e.g., regional hazard mapping or insurance portfolio analysis). This research develops a machine learning (ML) model that serves as a cost-effective proxy for traditional geotechnical testing.

Using CPT data from the New Zealand Geotechnical Database (NZGD), this study implements the state-of-practice Boulanger and Idriss (2016) methodology to calculate Liquefaction Potential Index (LPI) values for 5,879 unique locations across five Holocene geological units in New Zealand (i.e., windblown, human-made, estuary, river, and swamp deposits). ML models were trained separately for each geological unit to predict CPT-derived LPI, using three primary features: earthquake magnitude (M_w 5.0-8.0), peak ground acceleration (PGA) (0.05-1.2g), and groundwater table depth (0.5-15.0m). For each CPT location, the LPI was recomputed under sampled M_w-PGA-GWT combinations to create an expanded training set spanning plausible hazard and groundwater states. Using this training dataset, several ML methods were initially tested (i.e., gradient boosting, XGBoost, LightGBM, neural network, support vector machine), finally selecting LightGBM based on the best accuracy-training time trade-off. 

Model performance varied by geological unit: windblown deposits were captured well, achieving R²= 0.854, whereas river deposits reached only R²= 0.555, despite the latter having more training data. This finding demonstrates that depositional homogeneity, rather than data volume, can be more influential on ML performance in geotechnical applications. Feature importance analysis revealed balanced contributions to influencing predictions (i.e., magnitude: 33.7%, PGA: 34.5%, groundwater table depth: 31.8%), indicating the need to represent groundwater variability rather than treating shaking intensity as the sole dominant control. Validation against analytical LPI calculations for a synthetic scenario representing fully saturated conditions (M_w = 6.5, PGA = 0.4g, GWT = 0m) yields moderate agreement (R²= 0.491). The models tend to produce more conservative estimates for LPI < 5 and slightly underpredict for LPI > 40, likely reflecting systematic biases in the training data distribution, where extreme cases are underrepresented. Real-world application was also assessed by comparing predicted patterns with observed liquefaction manifestations during the 2011 Christchurch event from NZGD, independent of the training dataset. Comparisons observed good qualitative agreement with known high-susceptibility areas in eastern Christchurch, including zones near the Avon River and coastal margins. The proposed framework provides a scalable alternative to traditional CPT-based assessments, particularly for large-scale regional applications.