Majdi Flah.

Service-life tools use a power-law aging exponent to describe how chloride transport changes with age. Earlier studies established dependence on binder, exposure, reference time, and test protocol, while practice often exports a fitted exponent beyond the measured age range.

The stability of a mixture-specific exponent across calibration windows requires a separate check from regression fit quality. Diffusion, migration, ponding, simulated-marine, and field measurements also represent different transport quantities.

Refit early and full measurement windows within each mixture and protocol, compare their independent 100-year transport projections, and quantify divergence by protocol tier, early-age sampling, and binder class.

The contribution is a same-mixture, paired-window audit of the power-law export workflow. It separates the profile-diffusion cohort used for service-life inputs from a pooled cross-protocol summary, and distinguishes direct 90-day evidence from two-youngest-age fallback fits.

Evidence base

The measurement corpus contains 3,292 observations from 182 source studies and 1,285 mixture groups. The stability audit retains 312 trajectories from 44 studies; 79 trajectories use ASTM C1556 or NT Build 443 profile-diffusion evidence.

Collection

Published measurements were reconstructed from tables, figures, and text while preserving source, mixture, exposure age, protocol, and units. Same-age replicates were averaged before fitting.

Included

Each audited trajectory requires traceable mixture identity within one publication and protocol tier, positive transport values, and at least three distinct positive exposure ages. Direct early-window fits require at least two ages at or below 90 days; otherwise the two youngest ages form a labelled fallback.

Excluded

Untraceable mixtures, ambiguous exposure ages or conditions, unassignable protocols, RCPT-only evidence, and resistivity-only proxies fall outside the audit. Groups with one or two ages remain in the measurement inventory but cannot support the paired-window stability audit.

Ordinary least squares fits the logarithm of the transport response against logarithmic exposure age for each early and full window. Independent intercepts and slopes generate a 100-year projection ratio R100. The audit flags ratios of at least 2 or at most 0.5, then reports protocol-specific and sampling-specific rates, publication-cluster uncertainty, and sensitivity to the window and threshold.

Pairing fits within one source, mixture, and protocol limits source and mixture confounding. Independent fits retain changes in both intercept and slope. OLS supplies a common estimator where replicate-level variance is unavailable; the full-window fit supplies an available-data comparator.

High divergence affects 41.8% of profile-diffusion trajectories and 42.9% of the pooled corpus. The rate is 37.1% for 224 trajectories with direct early-window evidence and 58.0% for 88 fallback trajectories. Median full-window R² is 0.96, showing that fit quality and extrapolation stability measure different properties. Rates describe the audited evidence and the stated twofold, 100-year screen.

Before exporting an aging exponent into a service-life model, report its protocol, exposure-age basis, calibration windows, and projection sensitivity. The divergence ratio compares transport projections; converting it into an initiation-time multiplier requires a separate chloride-ingress calculation with defined cover, boundary conditions, and chloride threshold.

Fickian chloride models support service life calculations, and probabilistic approaches already account for uncertain exposure or corrosion thresholds. Machine learning studies estimate diffusion coefficients from heterogeneous experimental evidence.

The dispersion produced by diffusion calibration is commonly reduced to a point prediction. Its effect on threshold triggered maintenance decisions and their environmental and economic consequences requires explicit propagation.

Quantify calibration residual dispersion, propagate it into corrosion initiation distributions, and compare resulting intervention demand, greenhouse gas emissions, and discounted costs with a deterministic baseline.

The contribution is a calibration to decision chain that carries empirically estimated diffusion dispersion into finite horizon maintenance rules. It identifies when threshold crossings amplify parameter uncertainty into different intervention decisions.

Evidence base

141 experimental records from 33 journal articles, retained from 663 candidate records.

Collection

Published mixture and test observations were compiled. Cube strengths were converted to cylinder equivalents using a factor of 0.8.

Included

OPC concrete of normal weight with natural aggregates, ASTM C1556 or NT Build 443 bulk diffusion tests, and laboratory curing and exposure at 22 ± 3 °C. Superplasticizers were the permitted chemical admixture.

Excluded

SCMs, blended or alternative binders, recycled aggregates, and chemical admixtures outside the defined material domain were excluded. Electrical proxies and migration tests fall outside the bulk diffusion calibration domain.

Gradient boosting calibrates log10 Dapp. CTGAN augmentation is restricted to training. Residual dispersion is measured on 36 real test observations after duplicate pruning. A lognormal diffusion distribution and a chloride threshold distribution feed 15,000 Fickian simulations per case, followed by sealer and overlay policy scenarios, emissions accounting, and discounted costs.

Log space represents multiplicative diffusion scatter while retaining positive coefficients. Real test residuals provide the dispersion parameter. Keeping the transport model and decision rules explicit isolates how the same uncertainty changes engineering decisions.

The residual standard deviation is σlog = 0.162. The paper reports approximately 20–40% higher expected maintenance demand than deterministic estimates, with environmental and economic impacts reaching 1.5–2 times deterministic estimates within the reported 5th–95th percentile ranges. These results belong to the specified exposure and maintenance scenarios.

The framework supports inspection and maintenance planning through initiation percentiles, repair likelihood, intervention counts, and impact distributions. Application to a particular asset requires its exposure boundary conditions and repair policy; the analysis concerns corrosion initiation and stylized maintenance renewal.

Durability testing describes transport and deterioration, machine learning predicts material indicators, and LCA evaluates environmental burdens. Existing studies connect these domains using different performance measures and lifetime assumptions.

A predicted material indicator, an assumed design life, and an impact per cubic metre can represent different functions. Their coupling can therefore rank mixtures under incompatible exposure, deterioration, and maintenance assumptions.

Assess how chloride and carbonation evidence can enter service life models, examine the admissibility of machine learning outputs, evaluate functional units and maintenance assumptions in LCA, and define a shared comparison framework.

The deterioration informed sustainability framework uses a deterioration trajectory as the common representation across testing, prediction, service life, and environmental assessment. Its contribution is a set of compatibility rules for coupling these domains.

Evidence base

699 retrieved records led to 184 studies in the qualitative synthesis. A further 14 foundational works provide historical context outside the PRISMA workflow.

Collection

Web of Science was the primary search database, with coverage checks in Scopus, ProQuest, and Google Scholar. Title, abstract, and keyword searches covered four streams: durability experiments, durability prediction, durability informed LCA, and service life modelling.

Included

English peer reviewed journal articles and reviews on transport driven durability, service life, durability prediction, or LCA with explicit performance assumptions. Post 2019 work was prioritized, with earlier operative mechanistic formulations retained.

Excluded

Conference papers, theses, books, preprints, alkali activated systems, fibres, and geopolymers were outside the defined scope. Further exclusions covered duplicates, insufficient reporting, durability studies without quantitative transport parameters, mechanical only prediction, and environmental studies without life cycle inventories.

PRISMA screening was followed by structured extraction of mechanisms, test protocols, exposure, transport parameters, prediction targets and validation, service life formulations, functional units, boundaries, and maintenance scenarios. The synthesis evaluates compatibility between successive parts of the durability to LCA chain.

A cross domain synthesis can assess whether an output represents the state, exposure, and time evolution required by the next model. The shared deterioration representation supplies the conditions needed for interpretable environmental comparisons.

The review reports transport variation of one to two orders of magnitude across binder and test conditions. It identifies cases in which environmental rankings reverse when service life or transport informed functional units replace volumetric metrics. These are literature synthesis findings, with the framework defining how comparisons should be constructed.

The framework helps engineers compare binders under a common exposure, deterioration endpoint, functional unit, and maintenance scenario. It guides evidence selection for durability design and LCA; a mixture specific numerical ranking requires compatible inputs for the chosen application.

Vibration monitoring uses structural response measurements to identify changes in condition. Conventional workflows often extract modal or statistical features before applying a damage classifier.

Feature extraction and multiple sensor requirements can complicate monitoring. The study examines whether a network can learn damage information from raw acceleration while reducing the required measurement channels.

Develop BuildingNet, assess damage detection and localization, compare single and multiple channel inputs, and investigate network depth, training data size, and signal noise.

A one dimensional convolutional network performs feature learning directly on time domain acceleration. The contribution links this representation to classification and localization under a single channel measurement configuration.

Evidence base

Numerical structural cases and an experimental QUGS structure with 30 joints and 30 labelled damage classes involving loose bolts.

Collection

The published study uses acceleration histories. The 2020 companion thesis documents broadband excitation of a ten degree of freedom model, sampled at 1,000 Hz and segmented into signal windows.

Included

The companion numerical experiment contains a healthy state and ten single floor damage states, each defined by a 50% stiffness reduction. The final article adds experimental validation on QUGS.

Excluded

Simultaneous damage at multiple floors and mass changes are outside the companion numerical model. The final experimental acquisition protocol and sample selection require confirmation from the full article.

BuildingNet applies one dimensional convolutional feature extraction and classification to acceleration windows. Parametric studies vary network layers, training set size, channel count, and added noise. Numerical and experimental cases test damage detection and location prediction.

Temporal convolutions learn local signal patterns directly from acceleration. Comparing channel counts and noise conditions tests the feasibility of reducing instrumentation while retaining damage information.

The article reports successful localization in numerical mid rise structural cases and experimental QUGS validation. Gaussian noise levels up to 20% are investigated. The published record does not provide a verified single accuracy figure for this summary; thesis experiment scores are not presented as final article scores.

The work motivates monitoring with fewer measurement channels. Deployment requires validation for the asset’s geometry, sensor position, excitation, environmental variation, and plausible damage states. Single channel feasibility in the studied cases does not establish coverage of every damage scenario.

Civil structures were already monitored through sensor responses and visual records. Machine learning applications spanned bridges, buildings, dams, tunnels, and wind turbines.

The expanding literature used different inputs, algorithms, and damage targets. Engineers needed a structured account of their suitability and practical limitations.

Survey civil monitoring applications, compare algorithm strengths and weaknesses, and identify implementation requirements and research needs.

The review organizes image and vibration approaches around their inputs, learning tasks, and engineering outputs, making application differences explicit.

Evidence base

Published civil structural monitoring applications, summarized by structure, input, algorithm, and output. A retained study total is not verified.

Collection

The literature is assembled in comparative application tables. Exact search platforms, queries, and dates require confirmation from the final methods section.

Included

The verified scope is machine learning for civil damage detection using vibration measurements or images.

Excluded

Formal screening exclusions are not verified in the available sources. They are left unspecified pending the full journal methods.

The synthesis compares clustering, regression, and classification applications across the two monitoring streams. Application tables and critical discussion connect algorithms to input preparation and damage assessment.

Grouping studies by measurement type and task makes their operating requirements easier to compare. It supports selecting candidate methods for a defined monitoring problem.

The companion thesis discussion identifies CNN feature learning and transfer learning opportunities, data and computing demands, SVM kernel and scaling choices, and sensitivity to image quality. Research needs include linking visible damage to physical mechanisms and extending real time and unsupervised monitoring.

The review provides a starting point for choosing monitoring methods for aging infrastructure. Asset specific validation remains necessary because data quality, structure type, and damage definitions vary across applications.

Empirical strength models supported conventional mixture design. Modern composites introduced more complex relationships between constituents and compressive strength.

Prediction accuracy depends on the dataset and hyperparameters. Results from different modern mixtures therefore require careful interpretation before a model is selected.

Critically compare computational intelligence approaches for modern concrete strength and explain their parameters, performance, advantages, and limitations.

The review brings material specific applications together with model settings and reported accuracy, supporting a more informed comparison of candidate predictors.

Evidence base

153 initially collected publications, narrowed to 65 journal papers. These counts describe literature records.

Collection

Keyword searches covered strength, machine learning, and advanced concrete families. Search platform names and dates require full text confirmation.

Included

Compressive strength prediction for modern or novel concrete mixtures.

Excluded

Conventional concrete mixtures and predictions of other mechanical properties were outside the review’s defined scope.

The authors compare published algorithms, datasets, material systems, model parameters, hyperparameters, and reported accuracy. Each result is interpreted within its original study.

Strength is a continuous prediction target. Reviewing settings alongside material and data characteristics helps explain variation in regression performance.

The synthesis identifies the importance of data structure and model configuration. It supplies comparative guidance rather than a new common dataset benchmark or a pooled prediction error.

The findings help engineers identify model families worth validating for a specified modern mixture. Such validation can support strength estimation and mixture development under the project’s material and testing conditions.

Manual crack inspection requires time and can vary between inspectors. Earlier image methods detected cracks or classified orientations, while measurement depended on image processing and calibration.

A useful inspection workflow needs both a damage label and geometric measurements. Classification accuracy alone does not establish the accuracy of crack dimensions.

Train crack and orientation classifiers, combine their tasks, extract crack geometry, and compare image based measurements with field observations.

The workflow connects deep learning classification to geometric crack quantification. It evaluates the recognition and measurement tasks separately.

Evidence base

The companion thesis documents 10,000 images for binary classification, 4,800 for four orientations, and 6,000 for the combined five class model. The combined dataset contains 1,200 images per class.

Collection

Images were drawn from the Özgenel dataset, SDNET2018, Structural ImageNet, and online photographs. Ten field images from Western’s laboratory and Spencer Engineering Building were used for geometric verification.

Included

Close views, generally 25–50 cm from the surface, with a usable crack or intact concrete label. Images were converted to grayscale and resized to 227 × 227 pixels. Cohorts used a 60/20/20 training, validation, and test split.

Excluded

Manual quality screening removed blurred, distorted, low resolution images and cracks at image edges. A numerical quality threshold and exclusion count are not documented in the verified companion description.

Convolution, ReLU, pooling, dropout, and softmax layers classify crack presence and orientation. Image enhancement, an improved Otsu threshold, morphology, and geometric processing estimate length, maximum width, and direction. A physical scale converts pixels to dimensions.

CNNs learn visual features for classification. The geometric stage produces the dimensions needed for inspection, and its separate field comparison tests measurement performance.

The combined classifier achieved 97.63% training, 96.5% validation, and 96.17% test accuracy. Reported mean relative errors were 1.5% for length, 5% for width, and 2% for orientation; the companion field verification used ten images. Intersecting cracks and scale calibration limit transfer to new images.

The workflow supports concrete condition surveys by attaching dimensions and orientation to detected cracks. These observations can inform inspection priorities and engineering assessment; interpretation still depends on the member, loading, exposure, and calibrated image scale.

Empirical equations related concrete properties to composition and curing. Complex mixtures increased the challenge of representing these multivariable relationships.

Engineers needed guidance on flexible predictors and on the conditions supporting their reported performance.

Critically assess machine learning approaches for compressive, tensile, and shear strength and elastic modulus, and identify practical selection considerations.

The review links algorithm families and reported applications with recommendations based on the relationship being modelled and the data available.

Evidence base

Published prediction studies spanning four mechanical properties. A retained literature count is not verified.

Collection

A literature survey classified references by technique. Named search platforms, dates, and full queries require confirmation from the complete article.

Included

Commonly used, adequately described approaches to predicting the four properties in scope.

Excluded

The authors omitted redundant information and discounted insufficiently detailed or rarely used techniques to focus the review.

The critical synthesis compares neural networks, support vector machines, decision trees, and evolutionary approaches, including statistical performance measures and published comparisons with empirical equations.

The stated selection focuses on common methods with enough detail for assessment. Comparing them by property and data conditions supports practical model selection.

The review describes a non exhaustive set of comparisons using the same test data that favour machine learning over empirical equations. Recommendations relate method choice to nonlinearity, input count, and sample availability. A pooled accuracy estimate is not established.

The synthesis helps engineers evaluate candidate predictors for design inputs and mixture development. Reliable use requires testing under the intended material domain, property definition, and validation conditions.