*Important notice: This news reports on an unedited version of the paper which has been accepted. and is awaiting final editing. Scientific Reports sometimes publishes preliminary scientific reports that are not fully edited and, therefore, should not be regarded as conclusive or treated as established information.
This framework combines ensemble learning with earthquake-engineering features to prioritize unsafe reinforced concrete buildings, improving seismic screening after major earthquakes while reducing missed high-risk structures during emergency inspection response workflows.
Study: An engineering-informed voting ensemble framework for rapid seismic risk screening of reinforced concrete structures. Image Credit: PaNNa Pannnathorn/Shutterstock
A recent study published in Scientific Reports presents a machine learning framework developed for rapid seismic risk screening of reinforced concrete buildings after major earthquakes. The research focuses on the 2023 Kahramanmaras earthquake sequence (KMES) in Türkiye and introduces a safety-first ensemble learning model that combines algorithms with earthquake engineering principles. The study demonstrates that the framework can identify unsafe buildings with high reliability while minimizing the risk of classifying dangerous structures as safe.
Faster and Safer Seismic Risk Screening
Rapid seismic risk assessment has become essential in earthquake-prone urban regions following major seismic disasters. The February 2023 Kahramanmaras earthquakes in southeastern Türkiye caused widespread destruction across several cities, with thousands of reinforced-concrete buildings collapsing or sustaining severe structural damage. Emergency response teams faced major challenges because conventional assessment methods depend on extensive field inspections and expert engineering evaluations.
Many existing seismic vulnerability assessment methods and machine learning with thousands of reinforced-concrete buildings collapsing or sustaining models mainly focus on improving overall prediction accuracy. However, high accuracy alone does not guarantee reliable identification of unsafe buildings, particularly when hazardous structures constitute a small proportion of the dataset. Several earlier studies overlooked important engineering parameters such as local soil conditions, spatial variations in seismic demand, and the influence of evolving seismic design codes.
Researchers have worked to address the limitations by developing an engineering-informed, safety-first ensemble learning framework for rapid post-earthquake screening of reinforced-concrete residential buildings in the Nurdagi and Islahiye districts of Gaziantep. The researchers trained and validated the model using official post-earthquake damage assessment records provided by Türkiye’s Ministry of Environment, Urbanization, and Climate Change.
Engineering-Informed Machine Learning Framework
The researchers developed the framework using a large dataset containing information from 7,292 reinforced concrete buildings. The dataset included geographic coordinates, number of storeys, construction year, seismic demand parameters, and official post-earthquake damage classifications. Turkish authorities classified the buildings as either “Safe” or “Unsafe” using the Crisis Damage Assessment protocol.
Researchers incorporated peak ground acceleration and short-period spectral acceleration values derived from Türkiye’s national seismic hazard maps. They also included local soil conditions through Vs30 values, which represent the average shear-wave velocity in the upper soil layers. Using interpolation techniques, the team estimated seismic demand values for each building location.
The final feature set contained seven variables: building height, construction year, peak ground acceleration, spectral acceleration, Vs30, and two interaction terms connecting structural characteristics with seismic demand. These interaction terms were critical because seismic vulnerability depends on the combined influence of building age, structural height, and earthquake intensity, rather than on individual variables alone.
The framework employed a soft-voting ensemble architecture that combined three advanced gradient boosting algorithms: XGBoost, LightGBM, and CatBoost. The framework combined predictions from all three models through weighted probability averaging.
The researchers optimized the model using Türkiye’s TRUBA high-performance computing infrastructure. They also selected a lower decision threshold of 0.40 to create a more conservative classification system, ensuring that borderline structures would still receive further inspection.
Strong Performance in Post-Earthquake Screening
Testing on independent datasets showed that the proposed framework delivered strong predictive performance while maintaining a clear safety-first approach. The model achieved an unsafe-building recall score of 88.91%, identifying nearly nine out of 10 dangerous structures.
Saving this for later? Download a PDF here.
The researchers intentionally prioritized reducing false negatives over maximizing overall accuracy, despite achieving an overall accuracy of 65.73%. In earthquake risk mitigation, failing to identify an unsafe building can have far more serious consequences than generating additional inspections for safe structures.
One of the study’s most significant findings concerned high-rise reinforced-concrete buildings. The framework achieved a perfect 100% recall rate for buildings with six or more storeys, meaning the model successfully identified every unsafe high-rise structure in the testing dataset.
Mid-rise buildings also demonstrated strong performance, with a recall rate of 95.1%. These results highlight the framework’s effectiveness in detecting vulnerable multi-storey structures that often contribute to higher casualty risks during earthquakes.
The findings also reflected major structural patterns observed after the 2023 earthquakes. Older buildings constructed under outdated seismic regulations experienced substantially higher collapse rates. Mid-rise and high-rise reinforced concrete structures showed greater vulnerability in areas exposed to intense seismic demand and poor soil conditions.
Researchers introduced random noise into the dataset to simulate survey errors and GPS inaccuracies to evaluate reliability under real-world conditions during post-disaster assessments. The framework maintained stable predictive performance, demonstrating strong robustness in environments where data quality may be uncertain or incomplete.
The team also performed spatial transfer experiments to determine whether models trained in one district could generalize effectively to another district with different seismic characteristics. Although the framework maintained high recall in new geographic settings, false-positive rates increased noticeably. These results highlighted a key limitation of purely data-driven models when applied to regions with different building inventories, seismic conditions, and damage distributions.
Urban Resilience and Future Earthquake Response
The study shows that machine learning can strengthen rapid seismic risk assessment when combined with earthquake engineering principles and life-safety priorities. The proposed framework serves as a Tier-1 screening tool to help emergency authorities quickly identify high-risk buildings following major earthquakes. The model could reduce immediate inspection workloads by nearly 38%, allowing engineering teams to focus on the most vulnerable structures.
The framework also outperformed simple heuristic approaches such as the “Pre-1998 Rule” by incorporating seismic demand, structural geometry, and interaction effects into the prediction process.
Although the current system relies on precomputed hazard maps instead of real-time ShakeMaps, future versions may integrate live ground-motion data, richer structural parameters, and computer vision tools for automated building assessment. Overall, the framework provides a scalable and safety-oriented solution for improving post-earthquake emergency response and urban resilience planning.
Journal Reference
Pürsünlü, Ö., Çevik, A., et al. (2026). An engineering-informed voting ensemble framework for rapid seismic risk screening of reinforced concrete structures. Scientific Reports. DOI: 10.1038/S41598-026-52799-8, https://www.nature.com/articles/s41598-026-52799-8
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.