• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Wednesday, November 18, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Interpretable Machine Learning Models Predict Deterioration in Steel Box-Girder Road Bridges

by
November 18, 2026
in Technology
Reading Time: 6 mins read
0
Interpretable Machine Learning Models Predict Deterioration in Steel Box-Girder Road Bridges

Interpretable Machine Learning Models Predict Deterioration in Steel Box-Girder Road Bridges

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Researchers have developed an interpretable machine-learning framework designed to predict the future structural conditions of steel–concrete composite box-girder bridges by integrating basic bridge specifications with publicly accessible meteorological observations and highway traffic records. According to an investigation published in the International Journal of Concrete Structures and Materials by Gil Hwan Wang, Minsun Kim, Eunsoo Choi, and Jong-Su Jeon, the extreme gradient boosting algorithm, widely known as XGBoost, emerged as the most accurate model among nine evaluated machine-learning architectures. The primary motivation for this modeling initiative stems from the severe economic burden associated with maintaining expanding and aging transportation networks. While modern bridge inspection technologies can provide precise structural diagnostics, their substantial operational and financial costs severely limit scalability across entire national road inventories, generating a critical requirement for rapid, cost-effective computational tools that reliably estimate bridge damage indices from minimal, readily available parameters.

The engineering challenge addressed by the study reflects broader global infrastructure conditions where maintenance budgets struggle to keep pace with escalating demands for inspection, structural repair, and rehabilitation. In South Korea, where the empirical data were gathered, an increasing proportion of bridges have already exceeded thirty years of service life, with that share projected to rise substantially in the coming decades. Steel–concrete composite box-girder bridges constitute roughly sixty-three percent of the country’s national road bridge inventory because their high torsional stiffness and structural efficiency offer significant advantages in curved alignments. Unlike prestressed concrete structures, which predominantly experience cracking and material degradation, steel box-girder bridges feature steel girders whose primary degradation mechanisms involve corrosion, fatigue accumulation, and severe environmental exposure. Corrosion in these structural systems diminishes critical sectional properties, such as plate thickness and section modulus, thereby directly undermining load-carrying capacity and structural reliability.

To build a predictive dataset without requiring expensive bespoke field monitoring, the research team compiled and harmonized records from three official infrastructure and environmental platforms: the Facility Management System, the Korea Meteorological Administration, and the Traffic Monitoring System. The Facility Management System contributed standardized structural specifications and administrative records, including total bridge length, deck width, structure age, span counts, vertical clearances, and historical inspection evaluation scores. Environmental observations from sixty-six national meteorological monitoring stations provided thirty-year records of monthly precipitation and temperature fluctuations, from which freeze–thaw cycles and chloride exposure proxies were derived. Accumulated snowfall during the winter months of November through February served as an analytical proxy for de-icing salt application, while temperature shifts across the freezing point defined freeze–thaw frequencies. Operational traffic demand was characterized using annual average daily traffic records mapped directly to structural routes.

The target variable for the supervised learning models was the bridge system-level damage index, calculated under national safety inspection and diagnosis guidelines. This index aggregates weighted component-level defect ratings spanning both primary structural elements, such as decks, girders, substructures, bearings, and cables, and secondary elements, including expansion joints, drainage systems, pavements, and railings. Component condition assessments evaluate specific structural defects, such as cracking, spalling, corrosion, deformation, and leakage, translating quantitative measurements and inspector evaluations into condition grades ranging from A to E. The researchers assembled 1,066 valid inspection-transition pairs from 194 distinct steel box-girder bridges, requiring that every bridge possess at least three formal in-depth inspection records. This selection filter ensured that multiple distinct deterioration intervals were available for each structure, enabling the computational models to capture genuine longitudinal degradation trajectories rather than isolated, single-interval transitions.

The compiled database of 1,066 transition pairs across eleven structural, environmental, and operational variables was processed through data cleaning and min–max feature normalization before being split into an eighty percent training dataset and a twenty percent unseen holdout testing dataset. The investigators evaluated nine separate machine-learning algorithms: ordinary linear regression, ridge regression, k-nearest neighbors, decision trees, random forest, extreme gradient boosting, light gradient boosting machine, categorical boosting, and Fourier neural operator architectures. Hyperparameter optimization for the tunable models was executed through Optuna utilizing a tree-structured Parzen estimator algorithm and a median pruner across 60,000 trials or a two-hour cap. Optimization sought to maximize mean validation explanatory performance using five-fold cross-validation with shuffling, maintaining a fixed random seed of 42 to ensure objective reproducibility across candidate architectures.

Evaluation of the models on the holdout testing set demonstrated distinct variations in predictive precision and generalization capacity. Although the random forest model achieved the highest coefficient of determination on the training set at 0.996 and the lowest training root mean square error at 0.003, its testing coefficient dropped to 0.692, revealing notable overfitting to training points. In contrast, XGBoost attained the highest testing accuracy among all evaluated algorithms, delivering an R-squared of 0.804 and a root mean square error of 0.021. The narrower gap between training and testing metrics observed in XGBoost was attributed to its regularized boosting objective, tree pruning, shrinkage, and subsampling mechanics, which effectively reduced sensitivity to extraneous noise within the inspection training data.

To rigorously verify these comparative results beyond a single data partition, the authors conducted formal statistical benchmarking utilizing the Statistical Tests for Algorithm Comparison platform. The non-parametric Friedman test ranked CatBoost as the most stable model across cross-validation runs, closely followed by LightGBM, whereas holdout testing had favored XGBoost. However, subsequent post-hoc analysis using Holm test adjustments and CatBoost as the control method revealed no statistically significant differences among CatBoost, LightGBM, XGBoost, and random forest. Additional five-fold cross-validation among the three boosting models yielded average R-squared values of 0.720 for XGBoost, 0.714 for LightGBM, and 0.673 for CatBoost. Because XGBoost delivered the peak performance on the holdout partition alongside competitive cross-validation stability, the researchers selected it as the representative model, while noting that its performance represents a competitive boosting class rather than absolute statistical superiority over all alternatives.

To overcome the black-box nature of the selected ensemble tree model, the researchers implemented a dual interpretability architecture combining Shapley additive explanations, known as SHAP, with iBreakDown. SHAP calculates additive feature attribution based on cooperative game theory, providing a globally consistent measure of how each input parameter pushes a predicted damage score upward or downward relative to the baseline expectation. Global SHAP summary visualizations revealed that time-dependent factors overwhelmingly dominated future deterioration estimates. The existing damage index, the inspection interval duration, and structure age ranked as the three most influential predictors. High initial damage, extended time gaps between consecutive evaluations, and advancing structural age consistently exhibited strong positive correlations with elevated future damage scores, affirming the model’s physical fidelity in capturing cumulative structural wear.

In contrast to time-dependent metrics, static geometric parameters, including total bridge length, deck width, vertical clearance, and span counts, displayed minimal individual importance and lacked consistent directional influence in standard SHAP summary evaluations. Environmental and traffic parameters exhibited moderate individual contributions, where higher freeze–thaw cycles, elevated chloride exposure proxies, and heavier annual average daily traffic volumes shifted predicted damage values upward. However, because standard SHAP attributions assume local additivity, they can fail to reveal nonadditive variable interactions. The authors therefore incorporated iBreakDown, which decomposes predictive contributions sequentially to identify explicit pairwise interactions. This secondary analysis demonstrated that geometric properties act as critical latent amplifiers of physical damage rather than passive dimensional markers.

Global frequency analysis conducted via iBreakDown established that the interaction between annual average daily traffic and the number of bridge spans was the single most frequent dependency across the dataset. This interaction reflects the physical reality that repeated cyclic traffic loadings induce complex dynamic stresses and fatigue accumulation that vary sharply depending on continuous multi-span load distribution. The next most prevalent dependencies paired chloride exposure with geometric dimensions, specifically deck width, bridge length, and span counts. Complementary SHAP dependence plots confirmed that while narrower bridge decks exhibited minimal sensitivity to chloride levels, wider decks subjected to high chloride exposure displayed pronounced positive shifts in predicted damage. The physical surface area of wider decks dramatically magnifies salt accumulation and chloride ingress, accelerating corrosion in underlying steel components.

The practical engineering utility of the interpretable framework was demonstrated on three representative bridge samples withheld from model training, each representing distinct structural configurations and service histories. For Bridge A, a twenty-seven-year-old structure with an earlier damage score of 0.237, the model forecast a future damage index of 0.28, corresponding to elevated maintenance risk driven primarily by age, historical condition, and geometry, which outweighed lower freeze–thaw exposure. Conversely, Bridge B, possessing a similar age but larger dimensional clearances and a moderate four-year inspection interval, achieved a lower predicted score of 0.24 despite enduring severe freeze–thaw cycles.

The third evaluation case, Bridge C, illustrated the vital role of interaction sensitivity in young infrastructure assets. Despite being only three years old, Bridge C generated an elevated future damage prediction of 0.26. Force-plot decomposition demonstrated that the protective benefit of the bridge’s youth was severely undermined by its exceptionally wide deck of 25.9 meters combined with an eight-year gap between inspections. In practical deployment, translating numerical outputs into standardized maintenance grades provides infrastructure managers with rapid screening capabilities to prioritize budgets, while underlying SHAP force plots supply clear justifications for operational interventions. The authors emphasized that the current framework was derived from bridges with at least three in-depth inspections located primarily within specific Korean administrative regions, meaning that adapting the methodology to international inventories will require retraining on regional datasets that reflect differing environmental regimes, maintenance practices, and structural design standards.

Subject of Research: Interpretable machine-learning prediction of bridge deterioration

Article Title: Combining SHAP and iBreakDown to Interpret XGBoost-Based Future Condition Predictions of Steel–Concrete Composite Box-Girder Bridges

Article References: Wang, G. H., Kim, M., Choi, E., & Jeon, J.-S. (2026). Combining SHAP and iBreakDown to Interpret XGBoost-Based Future Condition Predictions of Steel–Concrete Composite Box-Girder Bridges. International Journal of Concrete Structures and Materials, 20(1), Article 91. https://doi.org/10.1186/s40069-026-00966-6

Image Credits: AI Generated

DOI: 10.1186/s40069-026-00966-6

Keywords: steel box-girder bridges, bridge deterioration, machine learning, XGBoost, SHAP, iBreakDown, structural health monitoring, infrastructure maintenance, damage index, civil engineering, freeze-thaw cycles, chloride exposure, traffic loading, explainable AI, Optuna, predictive modeling

News Source: Teresa Odom. (November 18, 2026). Interpretable Machine Learning Models Predict Deterioration in Steel Box-Girder Road Bridges. Scienmag.

Tags: bridge deteriorationchloride exposureCivil Engineeringdamage indexExplainable AIfreeze–thaw cyclesiBreakDowninfrastructure maintenanceMachine LearningOptunapredictive modelingSHAPsteel box-girder bridgesStructural Health Monitoringtraffic loadingXGBoost
Share12Tweet7Share2ShareShareShare1

Related Posts

Life Cycle Assessment Highlights Urban Waste Slurry Carbon Reduction Potential

Life Cycle Assessment Highlights Urban Waste Slurry Carbon Reduction Potential

November 18, 2026
Review Identifies Date Palm Seeds as Feasible Feedstock for Sustainable Biodiesel

Review Identifies Date Palm Seeds as Feasible Feedstock for Sustainable Biodiesel

November 18, 2026

Hybrid Model Minimizes Wire Electrical Discharge Machining Roughness in D2 Tool Steel

November 18, 2026

New Legal Framework Maps Accountability Across Artificial Intelligence Development Pipeline

November 18, 2026

POPULAR NEWS

  • Streaming Framework With Hybrid Machine Learning Achieves Sub-Forty-Millisecond Intrusion Detection

    Streaming Framework With Hybrid Machine Learning Achieves Sub-Forty-Millisecond Intrusion Detection

    29 shares
    Share 12 Tweet 7
  • Multidisciplinary Protocol Aids Mitral Valve Repair in Post-Infarction Shock Case

    29 shares
    Share 12 Tweet 7
  • Physical Exercise Linked to Lower Relative Health Disadvantage Among Older Adults

    29 shares
    Share 12 Tweet 7
  • Adolescent Brain Connectivity Predicts Depression and Anxiety Trajectories Over 18 Months

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Four Golgi-Related Genes Distinguish Endometrial Carcinoma From Normal Tissue in Model

Poly-Tobacco Use Rose Among Criminal Legal System Impacted Adults in Kentucky

Early Circumstances Drive Over Thirty-Four Percent of Adult Health Inequality in China

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.