Chloride-induced deterioration of reinforced concrete (RC) structures severely depreciates the mechanical advantages provided by the composite action of concrete and steel reinforcing. Degradation of the embedded steel and succedent effects on the surrounding concrete occur naturally over several decades, depending on the in-situ environmental conditions and construction quality. As many RC structures are reaching the end of their initial design lives, progressive deterioration of the reinforcing steel over time has created a far-reaching issue of compromised structural reliability, particularly in seismically active zones. Transverse reinforcement faces additional susceptibility to decay, due to its geometric and physical design relative to the longitudinal steel. The increased vulnerability alludes to an accelerated loss of shear and confinement-related mechanical properties as degradation advances with time. This Ph.D. research explores the implications of chloride-induced corrosion on the mechanical response of degrading RC members. The project is separated into two primary scopes. The first involves a comprehensive multi-phase experimental program to characterize the effect of corrosion on the cyclic shear degradation of affected circular RC columns. The second explores machine learning as a high-performing alternative to conventional modeling approaches, for predicting key mechanical properties in degrading RC members.

Artificially accelerated methodologies have become indispensable for achieving a broad spectrum of corrosion damages, enabling researchers to analyze the performance of affected reinforced concrete elements. However, an issue exists in over- accelerating the process, which creates unrealistic morphologies, limiting the experimental value of the test. An experimental study is undertaken to investigate the effects of the degree of acceleration on the corrosion morphology, sectional properties, and crack distribution of laboratory-scale RC elements. Twenty-four RC cylinders were constructed, corroded at two current densities, two concrete cover thicknesses, and using two variations of the impressed-current method. Results suggest a maximum upper bound threshold of 300 µA/cm2 should be used to ensure realistic corrosion morphologies and crack behavior. It is recommended that wet-dry phasing be implemented as the primary variant of the impressed-current method because of its superiority in producing morphologies, sectional properties, and cracking that adequately reflect those found in naturally corroding elements. This experimental phase also serves as a calibration benchmark for the proceeding artificial corrosion implementation on large-scale RC circular columns.

The ever-growing volume of experimental and field data continually enables advances in the field through deeper micro-macro analyses and various modeling applications. Machine learning offers powerful and robust methodologies for improving the predictive capabilities of complex phenomena, with the additional benefit of continual adaptability as more data emerges. An experimental phase is introduced, describing the tensile testing of 284 artificially corroded, 25 mm diameter deformed Grade500E reinforcing bars. Next, the mechanical characteristics of corroded bars are predicted through a collection of regression-based machine-learning algorithms. Models are trained and tested on a database of 1,387 tensile tests compiled from 25 other experimental programs available in the literature. The complete database includes 19 input parameters used to predict nine key mechanical properties of the corroded steel bars: yield force, ultimate force, effective yield stress, residual engineering yield stress (%), effective ultimate stress, residual engineering ultimate stress (%), ultimate strain, elastic modulus, and ductility. Nine machine learning models were selected from a balanced assortment of algorithm typologies to determine the most appropriate methodology for each response variable. Literature-supplied decay laws serve as null models for statistical analysis. Results indicate that machine learning applications can drastically improve the predictive accuracy for both strength and displacement-based characteristics. The adaptive-neuro fuzzy inference system (ANFIS) model was found to have the strongest individual predictive ability across all models. Meanwhile, ensemble tree-based learning algorithms categorically provided the most consistently high-performing models over the selected response variables.

Following on from the small-scale corrosion phase, a similar comparative assessment of two artificial corrosion techniques was undertaken in large-scale circular RC columns to quantify the authenticity of the induced damage patterns. The impressed-current method was divided into two subgroups – constant saturation and wet-dry phasing, at two imposed current densities of 200 µA/cm2 and 300 µA/cm2. The assessment was applied over a large-scale experimental program, testing the cyclic shear behavior of eight circular RC concrete piers. Analyses were conducted from the local corrosion distribution and morphology scale through to each test specimen’s global structural behavior and cyclic response.

The seismic shear response of fourteen scaled circular RC piers subjected to artificial chloride-induced corrosion is then investigated. Each pier was designed to trigger a shear-dominated failure and artificially corroded at two current densities. A further parametric study is undertaken, investigating the influence of confinement effective- ness, spiral diameter size effect, and pier aspect ratio on the rate of cyclic shear loss. As corrosion increases, a shift in failure mode towards a more brittle mechanism is consistently observed. Significant reductions in shear capacity were observed, with significant implications on the ultimate deflection capacity at severe levels of deterioration. In the most severe case, a maximum reduction of 37.5 % and 70.9 % were recorded in peak shear capacity and ultimate deflection, respectively.

A series of empirical modifications to an existing three-step analytical model is proposed to derive the cyclic shear capacity of circular RC columns considering corrosive conditions. The results of sixteen shear-critical RC columns, artificially corroded to various degrees and tested under quasi-static reversed cyclic loading, were used for model verification. Two simplified degradation models are developed to serve as performance benchmarks. The final model is proposed in a stepwise format relative to the measured damage of the steel reinforcement. New empirical decay law coefficients were derived for determining the degraded material properties based on the extensive database of over 1,380 corroded tensile tests. An additional experimental database of 45 corroded RC circular piers, with displacement ductilities ranging between 1.4 to 11.9, was collected to assist in the modification of ductility- based parameters. The results of the model, compared to experimental tests failing in shear-dominated modes, indicate that the peak shear capacity can be predicted well across a range of deterioration severities (0 – 58.5 % average transverse mass loss). As corrosion damage increases, the distribution of the corrosion relative to the location of the shear plane becomes a critical performance consideration, increasing predictive variance.

Lastly, an extensive database is introduced, aggregating 54 experimental programs with 804 test specimens and 45 input parameters, investigating the implications of chloride-induced corrosion on the deteriorated response of corroded reinforced concrete beams. Several statistical models are explored to determine the highest performing predictor for five key response variables – the residual ultimate moment capacity, residual capacity factor, yield load, yield displacement, and the ultimate displacement. Three existing analytical approaches were included for comparative purposes to test the efficacy of the trained statistical models. The optimized machine learning models outperformed conventional analytical approaches and achieved high levels of predictive accuracy. Ensemble tree-based learning algorithms consistently produced the best predictions.