Adaptive Kriging Framework Reduces Structural Reliability Evaluations by Twenty-Five Percent

Researchers have formulated an advanced structural reliability framework that merges Polynomial Chaos Expansion with adaptive Kriging modeling to overcome the computational bottleneck of high-dimensional engineering assessments. The methodology, designated as AK-DRPCE, incorporates targeted dimension reduction and an upgraded active learning function to evaluate failure probabilities with high precision while reducing costly performance function evaluations by up to twenty-five percent. Structural reliability analysis serves as an indispensable tool for determining whether civil infrastructure meets rigorous performance criteria under complex environmental and operational stresses. However, conventional surrogate models frequently falter when confronted with high dimensionality, complex nonlinear responses, or implicit performance functions that demand exhaustive finite element calculations. By deploying Polynomial Chaos Expansion to model overarching global trends and pairing it with a Gaussian process Kriging component to capture localized variations and predictive uncertainty, the new framework achieves a balanced mathematical architecture capable of navigating intricate limit states without succumbing to redundant sampling or computational exhaustion.

Over the past three decades, structural reliability engineering has predominantly relied on three overarching methodological pathways: approximation techniques, stochastic simulation algorithms, and surrogate modeling frameworks. Approximation methods perform adequately when evaluating elementary performance functions, but their foundational assumptions frequently precipitate severe convergence failures and compromised accuracy whenever applied to strongly nonlinear or high-dimensional systems. Stochastic simulation approaches, grounded firmly in the statistical law of large numbers, estimate failure probabilities by generating expansive sample populations and iteratively querying system performance; nevertheless, because practical structural failure events are governed by implicit mathematical relationships, repeatedly executing full-scale finite element models becomes computationally prohibitive. Consequently, surrogate models, including the Response Surface Method, Artificial Neural Networks, Support Vector Machines, standalone Kriging, and independent Polynomial Chaos Expansions, have emerged as crucial alternatives designed to construct surrogate input-output mappings from small sample designs. While Kriging provides unbiased predictions alongside rigorous estimation variances and Polynomial Chaos Expansion captures broad global behaviors, neither technique operating in isolation fully resolves multi-dimensional efficiency hurdles.

To exploit their complementary mathematical properties, engineers previously introduced combined PCE-Kriging models, wherein Polynomial Chaos Expansion serves as the trend function to establish global behavior while the underlying Gaussian process addresses residual discrepancies. In this unified framework, the performance function is initially approximated across the global parameter space, and the remaining residual variations are subsequently captured through a zero-mean Gaussian process governed by a calibrated covariance matrix. However, PCE-Kriging inherently inherits fundamental constraints associated with classical polynomial expansions, particularly the prerequisite that input random variables exhibit mutual independence. To circumvent correlated input variables in real-world structural systems, the researchers applied Nataf transformations combined with Cholesky decomposition, thereby mapping correlated random vectors into an orthogonal space of independent standard normal random variables while preserving essential correlation structures. Even with decorrelated standard variables established, increasing the number of input dimensions triggers an exponential escalation in the number of multivariate orthogonal basis polynomials, severely constraining standard PCE-Kriging implementations when analyzing multi-parameter structures.

The mathematical origin of this computational bottleneck lies in the combinatorial formulation governing multivariate orthogonal polynomial bases constructed through tensor products of univariate polynomials. For a system characterized by an input dimensionality of n and a total polynomial expansion order of d, the complete cardinality of multi-index combinations reaches (n plus d) factorial divided by the product of n factorial and d factorial. As the number of stochastic variables expands, the resulting basis set rapidly proliferates, precipitating the classical curse of dimensionality that destabilizes numerical estimations and consumes excessive computational resources. To mitigate this pervasive scaling challenge without resorting to unstable candidate-selection routines that degrade under uniform parameter importance, the AK-DRPCE methodology introduces a dedicated dimension reduction parameter denoted as s. By enforcing an explicit mathematical ceiling that restricts the maximum interaction order among participating random variables to s while preserving total orders up to d, the algorithm systematically eliminates unnecessary higher-order cross-interaction terms, significantly curtailing the expansion set while retaining robust expressive modeling capacity.

Beyond controlling basis dimensionality, the efficiency of adaptive surrogate modeling hinges fundamentally upon the formulation of the active learning function that governs sequential experimental point selection. Classical formulations, such as the Expected Feasibility Function and the probability-of-misclassification U-function, evaluate candidate points across the design space to locate regions with high predictive uncertainty near the limit state surface. Subsequent developments introduced composite learning functions that incorporated penalty terms for spatial sample distance alongside prediction variance; however, these historical approaches maintained static weighting coefficients throughout the entire iterative process. The AK-DRPCE framework addresses this limitation by enhancing the learning function into an adaptive formulation, designated as psi prime m, which incorporates a dynamically varying weighting coefficient alpha. Governed by the ratio between the average k-nearest neighbor distance in the current design of experiments and that of the initial experimental set, alpha increases progressively as iterations proceed from a calibrated lower bound of 0.2, dynamically shifting emphasis from initial spatial exploration toward focused local exploitation.

Because increasing the weighting coefficient alpha during advanced iterations reduces the penalization applied to sample proximity, standard adaptive schemes often suffer from sample clustering and spatial redundancy around the limit state boundary. To counteract this vulnerability, the researchers integrated an explicit local density control strategy grounded in the k-nearest neighbor principle into the psi prime m learning function. Whenever the algorithm identifies a candidate sample minimizing the learning metric, it evaluates the local geometric neighborhood surrounding that point relative to its three nearest neighbors within the existing design of experiments. The candidate point is approved for performance function evaluation only if its average neighbor distance matches or exceeds the minimum existing distance and the mean nearest-neighbor distance within the local cluster exceeds 1.5 times that minimum threshold. If a candidate point violates these spatial density criteria, the algorithm rejects it and sequentially inspects subsequent candidate points, thereby preventing redundant data clustering, enforcing uniform coverage, and conserving finite element evaluations.

The comprehensive operational architecture of AK-DRPCE follows a structured ten-step sequence designed to balance accuracy, exploration, and computational thrift throughout structural reliability assessment. The process commences by drawing an extensive candidate population from defined input distributions via Monte Carlo sampling, followed immediately by Nataf transformation into independent standard normal vectors. An initial design of experiments is established using Latin Hypercube Sampling, allowing the dimension-reduced PCE-Kriging surrogate to be constructed from baseline response evaluations. The surrogate model then estimates performance outcomes across the entire candidate population to determine initial failure probabilities via indicator functions. Model convergence is assessed across five consecutive iterations, requiring the average relative error of predicted failure probabilities to fall below a strict threshold of 0.001. If convergence criteria remain unsatisfied, the algorithm calculates psi prime m values, subjects the top candidates to local density screening, enriches the experimental design with the qualifying sample, and updates the surrogate; once converged, candidate pool sufficiency is verified by ensuring the coefficient of variation remains under five percent.

The computational accuracy and sample efficiency of the AK-DRPCE methodology were initially validated using a modified Rastrigin numerical benchmark characterized by severe nonlinearity and multiple disjoint failure regions. Operating across a two-dimensional stochastic input space with normally distributed variables exhibiting a correlation coefficient of 0.4, the algorithm was tested against fifty independently randomized initial designs of experiment. Comparative evaluations demonstrated that AK-DRPCE equipped with the adaptive psi prime m learning function achieved superior limit state curve fitting and captured disconnected failure domains that static-weight formulations completely missed due to localized sample clustering. Furthermore, the adaptive weighting parameter alpha successfully expanded from 0.25 to 0.68 across successive iterations, driving the algorithm toward rapid convergence. Across the fifty repeated trials, the method reduced the required number of performance function calls by twenty-five percent relative to standard psi m implementations and by thirteen percent compared to AKSE algorithms, while single-iteration computing time remained under one second on standard multi-core processing hardware.

To verify practical engineering applicability, the investigators deployed the AK-DRPCE framework to evaluate the ultimate limit state load-carrying capacity of a large-span cable-stayed bridge featuring a steel box girder. The physical infrastructure spans an aggregate length of 860 meters, configured with symmetric side spans of 215 meters and a central main span of 430 meters, alongside a 38-meter deck width, 137.5-meter towers, and 68 pairs of stay cables. Structural safety was formulated as an implicit limit state function where mid-span sectional stress induced by combined deadweight, traffic loading, and thermal gradients must not exceed a structural design resistance of 345 megapascals. The system was modeled within ANSYS and queried through a Python-based PyMAPDL interface, employing shell elements for the box girder, solid elements for piers and pylons, and link elements for stay cables across a nine-dimensional stochastic parameter space encompassing cable diameter variations, component elastic moduli, and dynamic environmental actions.

Stochastic vehicle and thermal actions were quantified using continuous field measurements gathered since December 2021 by the bridge’s structural health monitoring and weigh-in-motion systems. Traffic loading across a 100-year design reference period was simulated through a calibrated stochastic traffic flow model incorporating measured vehicle weights, time intervals, and an average speed of 19.4 meters per second. Thermal actions were established by analyzing full-year 2024 monitoring records, identifying peak temperature gradients on August 13 at 15:00 and modeling the joint statistical dependence of atmospheric temperature and wind speed through a bivariate Copula framework extrapolated to 100-year extremes. Reliability analyses executed across fifty randomized initial designs confirmed that AK-DRPCE with psi prime m reduced finite element calls by nineteen percent compared to standard psi m and seventeen percent compared to AKSE, with iteration times dominated by finite element analysis at three to four seconds. The researchers noted that future investigations will explore the coupled interactions between total polynomial order d and reduction parameter s across broader structural configurations.

Subject of Research: Adaptive surrogate modeling for structural reliability analysis

Article Title: An adaptive Kriging method with PCE-based dimension reduction for structural reliability analysis

Article References: Zhang, Z., Pu, Q., & Hong, Y. (2026). An adaptive Kriging method with PCE-based dimension reduction for structural reliability analysis. Advances in Bridge Engineering, 7(1), Article 45. https://doi.org/10.1186/s43251-026-00235-2

Image Credits: AI Generated

DOI: 10.1186/s43251-026-00235-2

Keywords: structural reliability analysis, adaptive Kriging, Polynomial Chaos Expansion, surrogate modeling, dimension reduction, finite element analysis, structural health monitoring, cable-stayed bridges, Monte Carlo simulation, active learning function, k-nearest neighbors, Latin Hypercube Sampling, weigh-in-motion, Copula modeling, Nataf transformation, thermal gradient analysis, ultimate limit state