Modeling Hydration Kinetics of Sustainable Cementitious Binders Using a Data Informed Nucleation and Growth Approach
Presented By: Taihao Han
Description: Predicting the hydration kinetics of [PC +SCM] binders is challenging for current analytical models due to the extensive diversity of chemical compositions and molecular structures present in both SCMs and PC, which result in a large number of independent parameters with intricate and nonlinear composition-performance correlations. This study develops an original deep forest (DF)-phase boundary nucleation and growth (pBNG) model to yield a priori predictions of hydration kinetics—i.e., time-resolved exothermic heat release profiles—of [PC +SCM] binders. The novel DF-pBNG model predicts time-dependent hydrate growth rate profiles for new [PC + SCM] based on their mixture design, and subsequently uses this information to reproduce their heat evolution profiles. This study utilizes a database that includes calorimetry profiles of 710 [PC + SCM] binder systems, encompassing a diverse range of commonly-used SCMs such as quartz, fly ash, limestone, among others, as well as both commercial and synthetic PCs. The results show that the DFpBNG model predicts the heat evolution profiles of [PC + SCM] in high-fidelity manner. Experimental results and outcomes of the DF-pBNG model are analyzed to rigorously evaluate the influences of SCMs on calorimetry profiles and growth rate of hydrates.
How to Optimize Concrete Deliveries Using Machine Learning and Concrete Truck On-Board Sensors
Presented By: Pierre Siccardi
Affiliation: Command Alkon
Description: To ensure proper concrete placement on site and targeted properties in the hardened state, ready-mixed concrete producers must be able to control the entire production cycle, from material batching to delivery. Weather conditions, intrinsic variability of raw materials, delivery conditions, or human factor are all parameters having a direct impact on the fluctuation of both fresh and hardened concrete properties, resulting in additional difficulties for producers. In order to further assist concrete producers, on-board sensor systems for concrete truck mixers have been developed over the last decade. By measuring the slump, the air content, the volume or the temperature of the fresh concrete inside the drum, those systems help ensure better concrete quality. It however comes at the cost of generating an ever-increasing amount of diverse data. This presentation will then expose how machine learning methods permitted to predict the concrete slump evolution during transportation from the plant to the construction site for a large number of delivery conditions.
Machine Learning Aided in Design, Characterization, and Monitoring of High-Performance Fiber Reinforced Cementitious Composite (HPFRCC)
Presented By: Weina Meng
Affiliation: Stevens Institute of Technology
Description: High-performance fiber reinforced cementitious composite (HPFRCC) features high compressive and flexural strength, strain hardening behavior, and long-term durability due to dense microstructure. The design of novel HPFRCC mostly depends on intensive experiments for trial and error, which is costly and inefficient. In addition, HPFRCCs exhibit dense microcracks whose opening widths are controlled within 100 µm. With narrow cracks, the cracked HPFRCCs have lower permeability and higher self-healing capability than conventional concrete. Thus, it is important to investigate the development process and the pattern of cracks in HPFRCCs, in order to understand the damages and degradation concerning the safety and durability. Conventional visual inspection and manual measurement for dense and microcracks is time-consuming and labor intensive. To accelerate the mixture development and characterization of HPFRCC, some advanced machine learning approaches have been developed: (1) Web clawer and table extraction methods were applied to extract mixture design information used to train machine learning models. (2) Machine learning methods were trained to predict key properties of HPFRCC. (3) Generative adversarial networks are used to generate cracked HPFRCC to enlarge the dataset. (4) Deep learning based semantic segmentation models are trained to convert the RGB image into binary image for crack identification and quantification.
Meta-Heuristic Optimization: Effective Machine Learning Techniques for Concrete Structures
Presented By: Nima Khodadadi
Affiliation: University of Miami
Description: A fundamental aspect of nature-based systems is the ability to optimize. Similarly, since ancient times humans have had a tenancy to naturally focus on optimizing their activities making them more feasible, economical, functional, and practical. Analogously in structural engineering design "structural optimization" is a simulation-driven design technique that identifies and explores high-potential designs, while also rejecting low- potential ones early in the design phase, aiming to solve problems of structural design. However, such optimization methods have not been widely used in the design of concrete structures. This is in part be due to the design and construction of concrete structures involves complex processes, and optimization techniques face serious challenges. Nevertheless, modern meta-heuristic methods of optimization, can provide higher-level procedures or heuristic designed to find, generate, tune, or select a heuristic (partial search algorithm) that may provide a sufficiently good solution to an optimization problem or a machine learning problem. In Machine Learning, historical data is used to teach and train the system developed, in order to be able to better predict future behavior. In meta-heuristic approaches, the need to have historical data is not necessary. Instead, the system generates random data and uses them to find an optimal solution that satisfies all the constraints. This iterative process continues until the algorithm reaches a defined criteria. Meta-heuristic algorithms are traditionally used in non-deterministic polynomial-time based problems, where for a given time and effort obtaining a “good” solution is preferred to an “optimum” one. To this end, meta-heuristic algorithms help select the optimal parameters for machine learning and deep learning techniques to train and improve the model's performance. Concrete structures optimization often aims to minimize costs, including those related to concrete material and reinforcement.
Exploring Machine Learning to Predict Concrete Field Performance Against Alkali-Aggregate Reaction (AAR)
Presented By: Ana Bergmann
Affiliation: University of Ottawa
Description: As one of the more harmful deterioration mechanisms affecting concrete infrastructures worldwide, the alkali-aggregate reaction (AAR) has been reported in over 50 countries. Among the several testing methods developed in laboratories to assess aggregate reactivity and the effectiveness of supplementary cementitious materials (SCMs) in mitigating AAR, the accelerated mortar bar test (AMBT) and the concrete prism test (CPT) are the most used around the globe. Moreover, field studies have been extensively developed to correlate laboratory tests with structures exposed to a real environment. Yet, current outcomes show significant discrepancies involving the mentioned laboratory tests, indicating no clear thresholds regarding aggregate reactivity potential for new structures. Nevertheless, although extensive work has explored the diagnosis of AAR on existing structures, there is still a lack of defining an accurate model for the prognosis stage. In this sense, the extensive current data on outdoor exposure sites requires implementing elaborated data analysis techniques (i.e., machine learning) to predict AAR development on both existing and new structures. Therefore, this work aims to explore how each variable affects AAR development through probabilistic approaches enhancing the accuracy of management protocols to assess the aggregate reactivity potential via laboratory tests to reduce the risks associated with AAR.
Vibration-Based Damage Detection of Continuous Structural Systems Using Physics-Informed Neural Networks
Presented By: Rui Zhang
Affiliation: Pennsylvania State University
Description: In this research, a physics-informed neural networks (PINNs) framework is developed for the damage detection of continuous structural systems based on limited and noisy sensor data. The PINNs framework integrates sensor data and knowledge of the physics of the system by embedding the sensor data, governing partial differential equations, and boundary conditions into the loss function of the neural network (NN) architecture. Through minimizing the physics-informed loss function, the NNs parameters and unknown structural parameters can be estimated, and hence the full state of the system is then predicted by the trained PINNs. The PINNs framework is demonstrated through application for the damage detection of a three-span continuous concrete beam subject to a dynamic moving load. The goal of the application is to detect the damage, which was modeled as the reduction of the flexural rigidity of each span of the concrete beam, as well as to predict the full state of the concrete beam. It is observed that the PINNs framework can accurately detect the damage and estimate the structural state from limited and noisy sensor data.