Fracture is one of the dominant deterioration and failure mechanisms affecting the serviceability and safety of concrete infrastructure. In practice, unavoidable variations in material properties and geometric configurations introduce significant uncertainties into fracture-governed structural responses, leading to substantial variability in structural performance and failure risk. Reliable assessment of such responses under uncertainty is essential for performance-based design and risk-informed asset management. However, the inherent complexity of fracture behaviour and the high computational demand of uncertainty quantification make such assessments particularly challenging. To address this challenge, this study proposes a phase field-based framework for stochastic assessment of fracture-governed structural responses in three-dimensional (3D) concrete structures. The framework combines a phase field model with an extended support vector regression (X-SVR) algorithm to establish mappings between uncertain inputs and targeted structural responses. To further improve predictive capability, a novel automatic relevance determination-generalised Hermite kernel (ARD-GHK) is developed and embedded into the X-SVR formulation. The resulting surrogate models enable rapid uncertainty quantification and statistical characterisation of fracture-governed structural responses over large ensembles of random realisations. Three representative 3D examples are employed to validate the proposed framework. The results demonstrate that the framework achieves excellent agreement with benchmark Monte Carlo simulations while reducing computational cost by several orders of magnitude. The proposed ARD-GHK is also shown to outperform polynomial and Gaussian kernels in surrogate modelling. By integrating high-fidelity fracture simulation with data-driven surrogate modelling, the framework provides an efficient computational approach for uncertainty quantification and structural safety evaluation of concrete infrastructure.
This study proposes an explainable artificial intelligence (AI)-guided framework for predicting compressive strength (CS) and assessing the sustainability of ultra-high-performance concrete (UHPC) mixtures incorporating supplementary cementitious materials (SCMs), nano-additives, and sustainable waste-derived constituents. A database containing 1990 experimental records of 28-day CS was compiled from the literature. Three tree-based machine learning algorithms, namely Extra Trees Regressor (ETR), Extreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM), together with stacked ensemble models, were developed and optimized using the Whale Optimization Algorithm (WOA). Based on outcomes, among all developed models, the SM-4 (XGB+ETR+LGBM) model achieved the best overall testing performance with an R² of 0.885, RMSE of 12.40 MPa, MAE of 8.80 MPa, MAPE of 5.94%, and RMSRE of 8.37%. SHAP analyses identified curing temperature, fiber dosage, silica fume content, cement dosage, and water content as the most influential predictors. Cradle-to-gate life cycle assessment (LCA)-based analyses further evaluated carbon emissions, embodied energy, and material cost. Additionally, a multi-objective inverse-design framework coupled the best-performing stacked model with NSGA-II and TOPSIS to identify UHPC mixture proportions for eight CS targets from 130 to 200 MPa while considering sustainability indicators and mixture-design requirements. Selected mixtures were produced and experimentally tested to compare measured CS with predictions and prescribed targets. Finally, an AI-guided graphical user interface (GUI) was developed for rapid CS prediction and sustainability assessment. The proposed framework integrates interpretable prediction, sustainability assessment, and multi-objective optimization to identify UHPC mixtures for prescribed strength targets, with experimental validation supporting practical mixture-design applications.
All-solid-state batteries (ASSBs) offer improved safety and energy density compared to conventional liquid–electrolyte systems, yet their widespread adoption is hindered by complex electro-chemo-mechanical degradation mechanisms. In this work, we introduce a stochastic continuum framework to capture microstructural heterogeneity and its influence on interfacial degradation in ASSBs. The approach combines physics-based modeling of electro-chemo-mechanical coupling, including a traction–separation law for interface decohesion, with non-Gaussian, spatially correlated random fields representing the variability of solid electrolyte properties. Random fields are constructed via a Gaussian translation method with Gamma marginals and a copula-based correlation structure, enabling realistic representation of particle-scale property fluctuations while maintaining computational efficiency. Numerical simulations on a representative half-cell of NMC811 cathode particles embedded in LPSCl electrolyte demonstrate the emergence of localized stress concentrations and interface damage under cyclic charging. The model quantitatively captures the evolution of specific differential capacity and depth of discharge, revealing how heterogeneous electrolyte properties amplify interfacial failure and drive catastrophic degradation. This framework provides a versatile tool for predictive design and reliability assessment of ASSBs accounting for microstructural variability.
For gravity dams constructed on soft rock, displacements at the same reservoir level may differ between filling and drawdown. Conventional inversion frameworks use the current reservoir level as the hydraulic input but neglect the loading path. Cyclic reservoir effects may therefore be misattributed to foundation creep, biasing the identified viscoelastic parameters. A reservoir-path-informed inverse framework is proposed to identify Burgers parameters from displacements at multiple monitoring points. First, an adaptive-window descriptor encodes reservoir-operation history using local water-level change, phase duration, and loading or unloading trend. Second, a physics-guided multi-output temporal convolutional network with factor attention (PG-ATCN) replaces repeated finite-element analyses during inversion. Separate branches encode viscoelastic-parameter, hydraulic-loading, creep-time, and reservoir-path information. Factor attention provides relative weights for selected physical factors, while a correlation-based output constraint preserves inter-point dependence and point-specific responses. Third, an adaptive multi-objective grey wolf optimizer (A-MOGWO) treats pointwise fitting errors as separate objectives, avoiding prescribed sensor weights. Archive-feedback operator selection, density-aware archive maintenance, and stagnation-triggered local refinement improve convergence, Pareto coverage, and error balance. Benchmarks against MOGWO and NSGA-II demonstrate A-MOGWO's performance. For a monitored roller-compacted concrete gravity dam, PG-ATCN reduced test-set RMSE by 22.8–35.5% relative to the best-performing baseline. Ablation analyses confirmed the contribution of reservoir-path information, and the identified parameters reproduced calibration and validation displacements at all monitoring points. The framework provides an approach to multi-point viscoelastic parameter identification in dam-foundation systems with path-dependent responses.
