Concrete Mix Design Optimization: Leveraging Machine Learning and Bayesian Optimization for Developing Low-CO2 Cost-Efficient Mixtures Containing SCM
Presented By: Arslan Akbar
Affiliation: Pennsylvania State University
Description: Advancements in AI and computational models have significantly enhanced the predictability of concrete performance by leveraging extensive datasets. Recently, machine learning models have been developed to predict concrete’s compressive strength based on its mixture proportions. However, these models treat supplementary cementitious materials (SCMs) as a categorical (as opposed to quantitative) parameter and do not account for the significant impact of the SCM reactivity on concrete’s strength development. In this study, we assembled a dataset of binary (cement-SCM) mixtures, incorporating SCM reactivity measured by the R3 (ASTM C1897) test. Utilizing a random forest machine learning model, we demonstrated that integrating SCM reactivity significantly enhances the model's predictive performance with the fewest input parameters (w/cm, SCM/cm, SCM R3 heat, Agg/cm, cement CaO%). Further, we implemented a multi-objective Bayesian optimization framework to assist in the mixture proportioning of low-carbon low-cost concrete utilizing cement(s) and SCM(s) available to a concrete producer. This framework proposes concrete mix designs to meet a target 28-day compressive strength while minimizing cost and CO2 emissions, by leveraging SCMs with varied reactivity levels. The proposed mix designs were further validated with experiments. The work demonstrates how to avoid model extrapolation and erroneous predictions by utilizing a multi-dimensional convex envelop algorithm. Overall, the outcomes of this work provide a valuable tool for the concrete industry which can be expanded to predict and incorporate other metrics of concrete performance (e.g., workability, durability) and develop optimized mix designs accordingly.
Carbon Assessment in the Proportioning of 3D Printable Concrete Mixtures
Presented By: Claudiane Ouellet-Plamondon
Affiliation: ETS Montreal, Universite Du Quebec
Description: The embodied carbon of concrete mixtures presents significant challenges, particularly for advanced applications such as sustainable 3D printing. Mixtures formulated for 3D printing often exhibit higher carbon footprints due to their specific performance requirements. This presentation focuses on adding carbon accounting to concrete mixture proportioning criteria. Two algorithmic approaches are explained. In one approach, we added a carbon criterion from a life cycle assessment to a multicriteria optimization genetic algorithm. In the other approach, we integrate carbon accounting into the proportioning criteria for 3D printable concrete. Our methodology builds on the CPIM (Compaction-interaction Packing Model) model. We employ a systematic experimental procedure that maps the compactness, cost, water-to-binder ratio, and carbon footprint of concrete mixtures on ternary diagrams, based on the proportions of constituents. This mapping allows us to identify regions of interest where performance and sustainability intersect. Experimental validation includes tests on selected mixtures to assess workability (free flow) and compressive strength. By correlating these properties with carbon metrics, we can pinpoint optimal formulations that minimize the environmental impact without compromising mechanical performance. Our findings offer a pathway to reducing the embodied carbon of 3D printed concrete, contributing to the broader goal of sustainable construction practices. This work highlights the importance of integrating environmental considerations directly into mixture design protocols, driving innovation in low-carbon, high-performance building materials.
An AI-informed System to Design and Inline Govern the Printability of Cementitious Composites
Presented By: Giacomo Rizzieri
Affiliation: Politecnico Di Milano
Description: The development of a printable concrete mix-design is a complex process that typically involves a trial-and-error phase for fine-tuning the exact proportions of each component. Predicting how rheological parameters will react to selected materials and dosages poses a significant challenge. In this paper, a methodology specifically tailored to regulate the printability of cementitious composites is introduced. The work starts with an experimental preliminary characterization of the rheological properties (via flow table test) of 3Dprintable mixes, taken from the literature. Then, leveraging a previously developed and calibrated numerical model (G. Rizzieri et al. 2023, Computation Mechanics), the approach has been employed to evaluate whether the concrete's static yield strength and viscosity align with key factors like printing speed, nozzle height, and the maximum number of admissible layers (all integral aspects of the printing process). Once the rheological characteristics are defined, a meticulously designed Artificial Neural Network (ANN) has been employed,. trained using diverse databases containing data on various concrete mixtures from existing literature, in order to correlate the rheological parameters allowing printability as per previously performed numerical modelling, to potentail mix-design contituents and dosages, ensuring they meet the specified requirements and the feasibility.
CNN-Based Approach for Evaluating Concrete Surface Integrity through Crack Identification and Analysis
Presented By: Majdi Flah
Affiliation: McMaster University
Description: Visual inspection remains the most common method for assessing concrete structures in service, where inspectors rely heavily on their experience, expertise, and engineering judgment to identify defects. Nevertheless, this approach is subjective, labor-intensive, and time-consuming, particularly given the difficulty of accessing many sections of complex structures. This study introduces a semi-automated inspection framework utilizing image processing techniques combined with deep learning to identify defects in concrete structures, especially in hard-to-access areas. The results demonstrate that integrating the Keras classifier with Otsu-based image processing achieves high classification accuracies of 97.63% for training data, 96.5% for validation data, and 96.17% for testing data. Additionally, the method exhibits minimal quantification errors of 1.5% for crack length, 5% for crack width, and 2% for crack orientation angle. Damage types and severity levels are classified according to permissible crack widths as defined by international standards and codes for various structures, including buildings and bridges. The proposed methodology is suitable for implementation using unmanned aerial vehicles, thereby providing a practical, nearly automated solution for inspecting and managing the extensive backlog of aging concrete infrastructure.