In a recent article published in the journal Results in Engineering, researchers proposed a novel deep-learning model to predict the compressive strength of slag-ash-based geopolymer concrete, an eco-friendly alternative to traditional cement. They also developed a software tool that suggests the best mix design for a given target strength and global warming potential.
Study: Predicting Eco-Friendly Concrete Strength with AI. Image Credit: shadesofquartz/Shutterstock
Background
Concrete is one of the most widely used construction materials globally, but it contributes to about 8% of global carbon dioxide emissions due to cement production, its main ingredient. To reduce concrete's environmental impact, practitioners are exploring alternative binders to replace or reduce cement use.
One alternative is geopolymer concrete, made from alkali-activated aluminosilicate materials like fly ash, slag, or rice husk ash. This concrete has a lower carbon footprint, higher durability, and superior strength to cement-based concrete.
However, geopolymer concrete has many variables affecting its performance, such as the type and proportion of the binder, the alkaline solution, the curing conditions, and the aggregates. Therefore, predicting the compressive strength of geopolymer concrete is challenging and requires extensive experimental work.
To address this, machine learning and deep learning models were used to learn the complex relationships between input variables and output strength. However, most of these models are black boxes that do not provide insight into their predictions, limiting their practical applicability and trustworthiness.
About the Research
In this paper, the authors developed a deep learning model to accurately predict the compressive strength of slag-ash-based geopolymer concrete. They utilized SHapley additive exPlanations (SHAP), a method that assigns a value to each input variable based on its impact on the prediction, considering interactions and dependencies among variables. SHAP provides global insights into the model, local explanations for specific instances, and feature dependency plots for individual variables.
The researchers analyzed a dataset comprising 260 samples of slag-ash-based geopolymer concrete with varied mix proportions and curing conditions. The dataset encompassed 11 variables: ground granulated blast furnace slag (GGBFS), corncob ash (CCA), fine aggregates (FA), coarse aggregates (CA), water (W), sodium hydroxide pellets (SHP), sodium silicate gel (SSG), curing days (CD), molarity concentrations (MC), concrete strength grade (CG), and compressive strength (CS). They split the dataset into training, validation, and testing subsets.
Additionally, the study compared three deep learning models: an artificial neural network (ANN), a deep neural network (DNN), and a one-dimensional convolutional neural network (1D CNN). The authors optimized model architectures and parameters using a grid search algorithm, assessing performance through errors and statistical indices. They applied SHAP to the most effective model to provide explanations for the predictions.
Research Findings
The outcomes showed that the DNN model surpassed the other models in accuracy and reliability. It achieved high coefficients of determination (R2) of 0.977, 0.982, and 0.972 for training, validation, and testing, respectively. The DNN model also exhibited the lowest mean absolute error (MAE), mean squared error (MSE), fractional bias (FB), and mean absolute percentage error (MAPE) compared to the other models. Its ability to capture intricate, non-linear data patterns enabled precise prediction of compressive strength in slag-ash-based geopolymer concrete.
SHAP analysis identified curing days as the most influential variable in predicting compressive strength, followed by corncob ash and ground granulated blast furnace slag. Longer curing periods and higher slag content were associated with increased strength, whereas higher ash content had the opposite effect. SHAP values showed interactions and dependencies among variables, such as a positive correlation between molarity concentrations and water content. Additionally, SHAP plots provided clear, intuitive visualizations of the model's decision-making process and the underlying reasoning behind its predictions.
Applications
The developed software tool employs an inverse modeling approach to predict mixed design proportions for a desired target compressive strength. It calculates the global warming potential (GWP) in kilograms of carbon dioxide equivalent for each mix design, ranking them based on their environmental impact.
Engineers and researchers can utilize this tool to optimize slag-ash-based geopolymer concrete for diverse applications such as buildings, bridges, pavements, and dams. It helps in selecting the optimal mix design that meets strength requirements while minimizing carbon emissions. The tool also facilitates an understanding of how various factors influence compressive strength and GWP.
Conclusion
In summary, deep learning and explainable artificial intelligence proved effective in explaining and predicting the strength of slag-ash-based geopolymer concrete. These methods offered accurate, reliable, and interpretable predictions, enhancing understanding and application within the construction industry.
The developed software tool represents a valuable asset for engineers and stakeholders, facilitating optimized mix design selection and promoting sustainable construction practices. Future work should expand the dataset to include more samples and parameters, such as varying types and sources of slag and ash, different alkaline solutions, diverse curing methods, and additional strength tests.
Deep Learning-based Gangue Sorting for Coal Plants