Can Large-Language (LLMs) Transform Concrete Material Science Across Various Types of Concrete?
Presented By: Yen-Fang Su
Affiliation: Louisiana State University
Description: Large language models (LLMs) such as GPT-4 have shown immense potential across a wide range of applications, from natural language processing to more unconventional domains like material science. Recent studies have suggested that these models provide value in chemistry and materials science research. To explore these possibilities, this study investigates the potential of LLMs to transform concrete material science, focusing on their applicability across diverse concrete types by capitalizing on their unparalleled capability to synthesize connections within massive bodies of textual data. We feed substantial literature on various advanced concrete types into cutting-edge LLM architecture. Specifically, three types of advanced concrete were investigated: ultra-high performance concrete (UHPC), self- healing concrete, and self-sensing concrete. The objective is to extract new patterns, correlations, and insights that are difficult to capture via a conventional machine-learning framework. By evaluating how LLMs can aid the development and modeling of these novel composites, this work provides insight into the future roles of AI in advancing concrete materials science. By combining the prolific knowledge generation of LLMs with first-principles modeling and experimental validation, LLMs could expedite discoveries and catalyze innovations of advanced multi-functional construction materials for the future.
Bridging Composition and Properties in Cementitious Materials with Machine Learning
Presented By: Kimberly Kurtis
Affiliation: Georgia Institute of Technology
Description: Predicting the properties of cement and concrete from composition remains a significant challenge due to the complexity and variability of the feedstocks, which leads to a high-dimensional problem with large degrees of uncertainty. Changes in cement compositions, particularly the increased blending with and/or co-grinding with limestone, along with expanded use of greater range of supplementary cementitious materials (e.g., natural pozzolans, reclaimed coal combustion residuals) have made the prediction of concrete properties significantly more challenging. This has resulted in additional risk in the design of concrete mixtures, structures and elements. This talk will summarize work on applying machine learning to the following challenges: (1) curating a comprehensive database describing the physical and chemical characteristics for cements currently in production across the US, (2) developing a data analytics-based model that links those characteristics to targeted performance criteria, and (3) recommending updates to design specifications to account for changes in cement characteristics.
On the Prediction of the Mechanical Properties of Limestone Calcined Clay Cement: A Random Forest Approach Tailored to Cement Chemistry
Presented By: Taihao Han
Affiliation: Missouri University of Science and Technology
Description: Limestone calcined clay cement (LC3) is a sustainable alternative to ordinary Portland cement, capable of reducing the binder’s carbon footprint by 40% while satisfying all key performance metrics. The inherent compositional heterogeneity in select components of LC3, combined with their convoluted chemical interactions, poses challenges to conventional analytical models when predicting mechanical properties. Although some studies have employed machine learning (ML) to predict the mechanical properties of LC3, many have overlooked the pivotal role of feature selection. Proper feature selection not only refines and simplifies the structure of ML models but also enhances these models’ prediction performance and interpretability. This research harnesses the power of the random forest (RF) model to predict the compressive strength of LC3. Three feature reduction methods—Pearson correlation, SHapley Additive exPlanations, and variable importance—are employed to analyze the influence of LC3 components and mixture design on compressive strength. Practical guidelines for utilizing these methods on cementitious materials are elucidated. Through the rigorous screening of insignificant variables from the database, the RF model conserves computational resources while also producing high-fidelity predictions. Additionally, a feature enhancement method is utilized, consolidating numerous input variables into a singular feature while feeding the RF model with richer information, resulting in a substantial improvement in prediction accuracy. Overall, this study provides a novel pathway to apply ML to LC3, emphasizing the need to tailor ML models to cement chemistry rather than employing them generically.
Real-Time Refinement of Concrete Mix Designs by Active Learning
Presented By: Mathieu Bauchy
Affiliation: University of California, Los Angeles
Description: Field production of concrete often deviates from its target performance as specified in the mix design, owing to variations in raw materials, environmental conditions, and batching protocols over time. These dynamic and intricate changes are challenging to quantify. As a result, these factors have posed a long-standing obstacle to achieving precision in concrete production, resulting in substantial material wastage, cost overhead, and environmental impact. While recent advances in artificial intelligence (AI) have enabled accurate prediction of various concrete properties, the utilization of AI to address variables in real production has seen limited success. Here, we introduce an innovative active learning scheme that enables on-the-fly AI model refinement, which is achieved by allowing the AI model to learn from the drifting of data patterns over time. This dynamic approach aligns the AI model with the ever-changing realities of concrete production environments, offering a robust and adaptable solution that bridges the divide between theoretical mix design and practical variations in large-scale concrete production.
Advancing Sustainable Construction: Predictive Modeling of Performance in Ultra-High-Performance Concrete(UHPC) through Machine Learning
Presented By: Hee-Jeong Kim
Affiliation: University of Arizona
Description: Ultra-high-performance concrete is one of the cement-based composite materials that has excellent strength and durability compared to conventional concrete. Typically, it contains cement, fine aggregate, silica fume, superplasticizer, and fibers. However, increasing the number of fibers significantly increases mechanical properties, compressive strength, and flexural strength but reduces workability.
In this study, we present selected machine-learning models to predict workability, compressive strength, and flexural strength as the output variables. The concrete
material incorporating PVA fibers and silica fume was adopted in this study. We performed some experimental tests to generate training data.
We evaluated three machine learning approaches and identified random Forest with fine-tuned hyperparameters as the best performing model that can help designing novel UHPC with minimal environmental impact.