A Hybrid Deep-Machine Learning Model with Bio-Inspired Optimization for Improved Soil Liquefaction Prediction

Soil liquefaction occurs when saturated soil loses strength due to excess pore pressure generated by seismic activity, often resulting in severe structural failures. Recent earthquakes have highlighted the need for accurate prediction and mitigation, especially in geotechnical engineering, where many interconnected parameters are difficult to define or model mathematically. Triggered by intense ground shaking, liquefaction can undermine the seismic response of urban infrastructure, making early prediction crucial for disaster resilience in densely populated areas. To address these challenges, Artificial Intelligence (AI) techniques—particularly machine learning (ML) and deep learning (DL)—offer a powerful alternative to traditional methods by effectively capturing complex, high-dimensional data patterns. In this study, we propose a hybrid framework combining the Jellyfish Search (JS) algorithm for hyperparameter optimization within an ensemble learning architecture. The model combines the feature-extraction capabilities of Convolutional Neural Networks (CNNs) with the classification performance of eXtreme Gradient Boosting (XGB). Data from Cone Penetration Tests (CPT) obtained from the literature are converted into image-like formats to leverage CNN capabilities before classification by XGB. Performance evaluations compared the proposed models against both standalone and hybrid models documented in previous studies. Among individual machine learning models, XGB outperformed others, followed by Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbors (kNN). The CNN model slightly exceeded existing standalone and hybrid ML-based models, including the Smart Firefly Algorithm with Least Squares SVM (SFA-LSSVM). When combined, the CNN-XGB model demonstrated superior predictive accuracy compared to either model used alone, highlighting the effectiveness of deep machine learning integration. The proposed JS-CNN-XGB model achieved the highest overall performance, with an additional 2.0% accuracy gain over the CNN-XGB model. These results indicate that XGB is the most robust predictive classification model, with CNN capturing complex features effectively, and that JS further enhances overall performance. Collectively, the JS-CNN-XGB model provides accurate and generalized predictions of liquefaction. Designed for civil engineers and construction risk managers, the system—featuring an embedded JS-CNN-XGB model—offers an intuitive interface and reliable analytical tools, functioning as a practical decision-support system for liquefaction risk assessment. Overall, these contributions emphasize the importance of integrating bio-inspired optimization with deep machine learning to address complex geotechnical challenges and turn research into practical solutions.