Active Learning Optimization Cuts Bridge Girder Computational Cost by Nearly Half

Civil engineers evaluating aging transportation infrastructure must contend with complex physical processes that steadily degrade load-bearing components over many decades. In a study published in Advances in Bridge Engineering, researchers Najib Zemed, Hanane Moulay Abdelali, Kaoutar Mouzoun, Toufik Cherradi, and Azzeddine Bouyahyaoui presented an active learning computational methodology designed to optimize the structural dimensions of reinforced concrete bridge girders over a full one-hundred-year operational lifespan. The investigation focused on resolving the steep computational barriers that have historically prevented transportation engineers from executing full probabilistic structural assessments. Traditional safety-factor methodologies typically apply uniform conservative margins that overlook how physical damage mechanisms interact dynamically over extended operational lifespans. Meanwhile, conventional stochastic simulation methods such as Monte Carlo sampling require exorbitant computational runtimes, often demanding tens of thousands of structural model calls that become practically intractable for routine civil infrastructure design workflows. To overcome these computational bottlenecks, the authors developed a two-stage reliability-based design optimization framework that couples deterministic baseline sizing with an adaptive hybrid surrogate modeling routine, substantially reducing required simulations while preserving strict long-term safety margins.

The engineering challenge addressed by the research team stems from the severe long-term operational environments sustained by highway bridges, where structures endure compounding degradation mechanisms rather than single isolated distress events. The authors focused their structural analysis on an actual highway bridge layout spanning twenty meters with an overall deck width of ten meters, comprising two three-and-a-half-meter-wide traffic lanes and two one-and-a-half-meter-wide pedestrian sidewalks. The superstructure consists of a two-hundred-millimeter-thick reinforced concrete deck slab simply supported across four regularly spaced T-shaped longitudinal girders, characterized by an effective flange width of two-and-a-half meters. Under classical structural design standards governed by ultimate and serviceability limit states without accounting for simultaneous long-term degradation interactions, the baseline girder geometry was sized at an initial depth of one point sixty meters and a web width of zero point fifty-five meters. The main tensile reinforcement within the bottom tension zone consists of ten thirty-two-millimeter diameter steel bars, supplemented by nine fourteen-millimeter diameter bars functioning as skin and assembly reinforcement, alongside web stirrups measuring ten millimeters in diameter for shear resistance and transverse slab rebar spaced at fifteen-centimeter intervals.

Rather than evaluating deterioration phenomena in isolation as seen in prior literature, the research framework integrated four interrelated time-dependent degradation and demand variables across the structure’s projected century of service: concrete creep, concrete shrinkage, chloride-induced reinforcement corrosion, and progressive traffic growth. To model vehicle loading authentically, the investigators moved beyond conventional static traffic rules by implementing empirical data captured via Weigh-In-Motion sensors at the Pontecagnano station in Italy, a facility chosen because its heavy commercial vehicle patterns reflect the high-density freight corridors linking Tangier-Med across Morocco into sub-Saharan trade routes with unrestricted gross vehicle weight limits. The dynamic axle distributions and spacings from this system were coupled with local inter-vehicle spacing records from Morocco’s Reggada region, where headway distance follows an exponential probability distribution representing free-flow traffic. The baseline highway corridor experiences an annual average daily traffic volume of approximately one thousand six hundred ninety-seven vehicles per day, subject to a documented annual compounding traffic growth rate of one percent that steadily escalates service stresses over time.

The physical degradation models chosen for the framework capture the progressive mechanical softening of the composite reinforced concrete cross-section over its century-long service window. Concrete drying shrinkage and sustained load creep compliance were formulated using the CEB-FIP 90 international model specifications, tracking the evolution of shrinkage strain and notional creep coefficients across multi-decade operational horizons. To model steel deterioration from aggressive environmental exposure, the authors incorporated the Stewart model, which establishes chloride penetration timing and calculates the progressive reduction of rebar diameter once chloride thresholds are breached. In parallel, mechanical concrete-steel interactions were captured through homogenized composite section properties, updating the neutral axis location, the uncracked moment of inertia, and the cracked moment of inertia as steel cross-sectional areas diminished. As tensile steel cross-sections corrode, the neutral axis shifts upward, widening tensile cracking in the concrete matrix and accelerating overall structural flexibility, which markedly amplifies mid-span flexural deflections under successive load cycles.

To rigorously predict time-dependent deflections without prohibitive numerical overhead, the investigators utilized a curvature-based analytical approximation grounded in classical beam theory, integrating elastic strain, cumulative creep curvature, and drying shrinkage curvature across the beam span. Because traffic loads grow continuously on an annual basis, creep deformations cannot be evaluated as a single sustained static load; instead, the researchers implemented a multi-pulse accumulation technique that discretizes annual moment increases and accumulates creep curvature increments based on the specific structural age at which each load increase occurs. Guyon-Massonnet semi-analytical transverse distribution theory was employed to calculate lateral load sharing across the four adjacent longitudinal girders under moving vehicular convoys, deriving torsional and flexural distribution parameters from girder and transverse diaphragm geometries. The maximum allowable serviceability deflection was set at the girder span divided by eight hundred, establishing the permissible operational boundary of zero point zero twenty-five meters in accordance with American Association of State Highway and Transportation Officials specifications.

The core computational innovation introduced by the authors centers on their two-stage Reliability-Based Design Optimization formulation, which balances initial deterministic exploration against high-precision probabilistic boundary refinement. The first stage executes a deterministic optimization using the Sequential Quadratic Programming algorithm, seeking to minimize the girder cross-sectional area while enforcing the deflection limit at one hundred years using mean variable values. This initial phase defines geometric limits bounding the aspect ratio between one point sixty-seven and three point thirty-three, while constraining web width between zero point forty-five and zero point eighty meters and girder height between one point three and two point one meters. Solving this deterministic formulation produced an optimal baseline configuration of zero point forty-six thirty-six meters in width and one point fifty-four thirty-eight meters in height within eleven iterations and only twenty-two finite element model calls, establishing an efficient starting point that dramatically accelerates subsequent probabilistic search loops.

The second stage refines this geometric baseline by applying a fully probabilistic reliability constraint that limits the structural failure probability to zero point zero sixty-six eight at one hundred years, corresponding to a target reliability index of beta equal to or greater than one point five as prescribed by Eurocode standards for serviceability limit states. Evaluating this limit-state function within repeated design iterations typically creates extreme computational bottlenecks. To resolve this, the authors constructed a weighted Ensemble Metamodel combining three diverse surrogate modeling architectures: Artificial Neural Networks, Radial Basis Functions, and Support Vector Regression. Model predictions are combined through error-weighted averaging using root mean square errors calculated from fivefold cross-validation. Furthermore, the variance across the three surrogate predictions serves as an explicit mathematical indicator of epistemic uncertainty, allowing the algorithm to quantify where the models disagree across the multidimensional input domain.

To minimize unnecessary structural analyses during surrogate training, the team embedded an active learning adaptive sampling strategy governed by the U-learning objective function. At each active learning step, the algorithm surveys a candidate Monte Carlo pool and identifies the specific sampling location that minimizes the ratio between the ensemble absolute prediction magnitude and its predictive standard deviation. This mathematical criterion systematically concentrates new training points in regions that are simultaneously close to the failure boundary and characterized by elevated model uncertainty, avoiding wasted simulations deep inside safe or clearly failed domains. Once the most informative point is identified, a single high-fidelity simulation is executed, the Design of Experiments dataset is updated, and all three surrogate models are retrained from scratch with updated hyperparameters, progressively sharpening limit-state prediction fidelity until strict convergence criteria are achieved.

Before optimizing the full bridge structure, the researchers validated their active learning ensemble surrogate methodology against a well-documented non-linear benchmark problem from structural reliability literature involving two random variables and three non-linear constraints. The benchmark results showed tight agreement with established published solutions across various target reliability indices, confirming the stability of the active learning ensemble framework. When tested on failure probability estimation for the bridge model, the proposed approach achieved a calculated failure probability of zero point zero forty-six nine using only seventy-nine finite element calls, yielding an error rate of just three point ninety-nine percent relative to a brute-force Monte Carlo baseline of one hundred thousand evaluations. In contrast, alternative surrogate formulations such as Adaptive Kriging Monte Carlo Simulation required one hundred nineteen calls with a five point thirty-two percent error, while Adaptive Kriging with Subset Simulation required ninety-five calls with a four point sixty-five percent error.

When applied to the comprehensive two-stage reliability-based optimization of the bridge girders, the framework achieved final convergence in fifteen probabilistic iterations following the eleven deterministic cycles, totaling two thousand and six finite element model evaluations. By comparison, an identical optimization performed using standard Adaptive Kriging Monte Carlo Simulation required nineteen iterations and three thousand eight hundred eighteen evaluations, demonstrating that the active learning hybrid ensemble achieved a forty-seven percent net reduction in computational cost. The final optimized girder dimensions converged to a web width of zero point forty-nine seventeen meters and a depth of one point sixty-three seventy-nine meters, establishing a cross-sectional area of zero point eighty fifty-four square meters that fulfilled the target reliability constraint throughout the century of simulated degradation. Sensitivity analyses further revealed that cumulative traffic growth and concrete creep coefficients were the primary drivers governing long-term performance degradation, demonstrating that infrastructure longevity depends heavily on accounting for coupled environmental and operational loading within computationally viable design workflows.

Subject of Research: Reliability-based design optimization of reinforced concrete bridge girders using active learning and hybrid surrogate modeling

Article Title: Optimization of reinforced concrete bridge girders using reliability-based design and active learning to ensure long-term serviceability

Article References: Zemed, N., Abdelali, H. M., Mouzoun, K., Cherradi, T., & Bouyahyaoui, A. (2025). Optimization of reinforced concrete bridge girders using reliability-based design and active learning to ensure long-term serviceability. Advances in Bridge Engineering, 6(1), Article 37. https://doi.org/10.1186/s43251-025-00184-2

Image Credits: AI Generated

DOI: 10.1186/s43251-025-00184-2

Keywords: Reliability-Based Design Optimization, Reinforced Concrete Bridges, Active Learning, Surrogate Modeling, Artificial Neural Networks, Radial Basis Functions, Support Vector Regression, Bridge Girder Optimization, Serviceability Limit State, Long-Term Deflection, Concrete Creep, Concrete Shrinkage, Chloride-Induced Corrosion, Traffic Growth, Sequential Quadratic Programming, Weigh-In-Motion, Finite Element Modeling, Structural Reliability