In recent years, the challenge of accurately assessing soil liquefaction potential during seismic events has remained a critical concern for geotechnical engineers and urban planners worldwide. Soil liquefaction, a phenomenon where saturated soil substantially loses strength and stiffness in response to earthquake shaking, poses significant risks to infrastructure and human lives. Traditionally, the evaluation of liquefaction susceptibility has relied on empirical correlations and simplified analytical models. However, these methods often lack transparency and may fail to capture the complex interactions governing soil behavior under dynamic loads. In a groundbreaking study published in Environmental Earth Sciences, researchers have unveiled a pioneering approach that leverages explainable artificial intelligence (XAI) frameworks combined with ensemble machine learning methods to revolutionize the estimation of safety factors against soil liquefaction.
At the core of this novel methodology is an integrated system employing SHapley Additive exPlanations (SHAP) with a Borda count-based ranking mechanism, united with ensemble machine learning algorithms. Ensemble learning, by aggregating multiple predictive models, enhances overall accuracy and robustness beyond what individual models typically achieve. The innovative fusion with SHAP explanations facilitates a transparent interpretation of model outputs, illuminating the influence of each input variable on final predictions. Such explainability is crucial when deploying AI-driven tools for critical infrastructure safety assessments, ensuring that engineers and decision-makers can understand, trust, and verify model recommendations.
The research team, spearheaded by Dağdeviren, Demir, and Erden, assembled an extensive database encompassing geotechnical parameters widely recognized as influencing liquefaction potential. These parameters include relative density, shear wave velocity, standard penetration test (SPT) blow counts, and other site-specific soil properties derived from seismic records and field investigations. The integration of diverse datasets was meticulously handled to train ensemble models capable of capturing non-linear relationships often encountered in subsurface soil conditions. Consequently, the AI approach discerns subtle patterns within data that conventional methods might overlook or misconstrue.
One breakthrough of this study lies in its emphasis on Explainable AI, particularly SHAP, which attributes the predicted output to individual feature contributions in a manner consistent with game theory. SHAP values allow practitioners to quantify the marginal effect of each input, providing insight into the decision-making process of black-box models such as Random Forests, Gradient Boosting Machines, and other ensemble techniques. This level of interpretation surpasses traditional “black box” constraints and enables rigorous scrutiny of model reliability, thereby bridging the gap between cutting-edge AI and practical engineering applications.
Moreover, the implementation of the Borda count algorithm innovatively addresses the challenge of reconciling feature importances derived from multiple ensemble members. By applying this voting-based ranking system, the model identifies and prioritizes the most influential geotechnical parameters affecting liquefaction safety factors. This ensures that the critical variables underpinning predictive outcomes are consistently recognized, decreasing the risk of model bias and enhancing the robustness of safety recommendations for seismic hazard mitigation.
From a practical standpoint, the research provides compelling evidence that ensemble machine learning with integrated SHAP-Borda methodology yields superior performance metrics over traditional empirical correlations. Model validation using a comprehensive test dataset demonstrated increased predictive accuracy in estimating the factor of safety against soil liquefaction, which is paramount for designing earthquake-resilient foundations and urban infrastructure. The ability to reliably quantify safety margins contributes directly to improved risk management strategies and cost-effective engineering solutions.
This approach also presents transformative implications for regulatory frameworks and decision support systems. By offering transparent and interpretable predictions, the AI model can serve as a trustworthy tool for geotechnical experts to complement or even challenge established design codes and guidelines. The enhanced explainability fosters collaboration and consensus-building among multidisciplinary stakeholders, including engineers, city planners, insurers, and emergency response teams, accelerating the integration of AI insights into practice.
The methodological architecture devised in this study involves a careful orchestration of data preprocessing, model training, and feature explanation phases. Raw input data underwent normalization and inconsistency checks to minimize noise and enhance model generalizability. Multiple ensemble algorithms were explored, including Random Forest, Extreme Gradient Boosting (XGBoost), and LightGBM, to optimize predictive accuracy and computational efficiency. Subsequently, the SHAP framework was applied to the best-performing model, unraveling the otherwise opaque decision boundaries into comprehensible feature impact assessments.
The scientific novelty also encompasses the harmonization of SHAP values using the Borda count, which aggregates rankings across ensemble components rather than relying on isolated single-model explanations. This consensus-driven approach minimizes overfitting risks and accounts for variability in feature importance distributions, ultimately culminating in a more reliable hierarchy of soil parameters influencing liquefaction potential. Such nuanced feature selection enhances interpretability while simultaneously facilitating model simplification without sacrificing accuracy.
Beyond immediate engineering applications, this integration of explainable ensemble AI techniques signals the broader promise of combining advanced machine learning with domain-specific knowledge in civil engineering and earth sciences. It showcases how interpretability frameworks can unlock hidden insights from complex datasets, fostering innovations that transcend traditional computational modeling boundaries. Future research can expand upon this foundation to incorporate time-dependent ground motion data, real-time monitoring inputs, and multi-hazard interactions, further enhancing predictive capabilities for seismic risk assessment.
Importantly, the study underscores an ethical dimension in AI deployment by advocating for transparent and accountable algorithms in areas where human safety is at stake. Explainability tools like SHAP provide a safeguard against unintended consequences stemming from misunderstood or misapplied AI outputs. This approach aligns with growing international calls for responsible AI adoption within infrastructure design, environmental engineering, and disaster resilience communities.
Furthermore, the adoption of ensemble machine learning combined with explainability techniques addresses long-standing limitations inherent in empirical and semi-empirical models traditionally used in geotechnical earthquake engineering. By circumventing restrictive assumptions and incorporating richer, multidimensional data representations, the proposed framework enables more nuanced and site-specific vulnerability assessments. This paradigm shift holds the potential to revise existing methodologies and standards deeply rooted in historical practice.
In practical workflows, the implementation details described in the research provide a reproducible protocol for practitioners seeking to harness explainable AI models. The researchers emphasize the integration of user-friendly computational tools and visualization dashboards to present SHAP-derived feature impacts interactively. Such accessibility ensures that professionals without advanced AI expertise can readily interpret model outputs and make informed decisions, strengthening interdisciplinary communication between data scientists and civil engineers.
Finally, the adoption of this explainable ensemble machine learning framework exemplifies the accelerating trend of AI-driven innovation addressing real-world infrastructure challenges. As urban centers expand and climate-driven hazards increasingly threaten built environments, robust and transparent risk estimation models become indispensable. The work by Dağdeviren and colleagues marks a significant milestone in this trajectory, offering a scientifically rigorous, interpretable, and performant solution to the persisting problem of soil liquefaction risk assessment in seismic regions.
Subject of Research: Estimation of the safety factor against soil liquefaction using explainable artificial intelligence and ensemble machine learning techniques.
Article Title: Explainable AI using ensemble machine learning with integrated SHapley additive explanations (SHAP)-Borda approach for estimation of the safety factor against soil liquefaction.
Article References:
Dağdeviren, U., Demir, A., Erden, C. et al. Explainable AI using ensemble machine learning with integrated SHapley additive explanations (SHAP)-Borda approach for estimation of the safety factor against soil liquefaction. Environ Earth Sci 84, 507 (2025). https://doi.org/10.1007/s12665-025-12466-z
Image Credits: AI Generated
