A study conducted by De-Cheng Fenget al., at Southeast University in Nanjing, China, focuses on
predicting the compressive strength of concrete. Compressive strength is a critical mechanical property of concrete, serving as an essential indicator of a structure’s strength, durability, and overall integrity. Traditionally, this parameter is determined through physical experiments involving the preparation of cube or cylinder specimens, which are then cured for a specific duration. The compressive strength is measured using a compressive testing instrument.
While some researchers have proposed empirical regression methods for this prediction, the problem’s inherent non-linearity often complicates the development of accurate regression expressions. Additionally, other researchers have explored numerical simulations; however, the complexities associated with non-linearity and random coupling present significant challenges in accurately replicating concrete behavior.
With the advancement of machine learning techniques, researchers worldwide have increasingly embraced various algorithms to predict the compressive strength of concrete. Some of the most prominent methods include Artificial Neural Networks (ANN), Support Vector Machines (SVM), heuristic regression methods, and Random Forests (RF). These machine learning approaches have consistently demonstrated superior performance compared to traditional predictive methods.
Studies indicate that ANN and SVM are among the most widely utilized techniques for predicting the compressive strength of concrete. ANN, with its ability to model complex, nonlinear relationships, allows for more accurate predictions by learning from a large set of data. Meanwhile, SVM is particularly effective in classification tasks and has shown great promise in regression applications as well, providing
robust predictions even in scenarios with limited data.
Overall, the integration of machine learning techniques into the field of concrete strength prediction has
significantly improved predictive accuracy, making it a vital area of research in civil engineering and materials science. As more researchers continue to explore and refine these methods, the potential for improved concrete performance and durability expands, benefiting construction practices and infrastructure development globally.
The author proposes using the Adaboost technique to predict the compressive strength of concrete, utilizing a dataset of 1,030 samples. The study employs a 10-fold cross-validation method to validate the developed model. Furthermore, an additional 103 samples were used to assess the model’s generalization capability. The performance of the developed model is compared with Support Vector Machine (SVM) and Artificial Neural Network (ANN) approaches. The input parameters for predicting compressive strength include: cement (kg/m³), water (kg/m³), coarse aggregate (kg/m³), fine aggregate (kg/m³), superplasticizer (kg/m³), blast furnace slag (kg/m³), fly ash (kg/m³), and curing time (days). The dataset was divided into 90% for training and 10% for testing.
The optimal hyperparameters for the AdaBoost model were determined using the grid search technique and are as follows: maximum iterations = 200, learning rate = 0.2, maximum depth below root = 50, minimum samples for a split = 5, minimum samples per leaf node = 2, and minimum impurity = 0.0001. The AdaBoost model was evaluated based on these optimal hyperparameters and compared with Support Vector Machine (SVM) and Artificial Neural Network (ANN) using the R-squared metric. The results
indicate that the R-squared value for AdaBoost is 0.982, while for ANN it is 0.903, and for SVM, it is 0.855. Thus, AdaBoost outperformed the other models.
Furthermore, the relative importance of input features was assessed using AdaBoost. The results revealed
that curing time is the most critical factor influencing final compressive strength, followed by cement, which has a significant impact on compressive strength (about 97% as influential as curing time). Cement plays a vital role in binding all other components together through the hydration reaction. Water ranked third, accounting for 57% of the influence of curing time, as it is essential for the hydration of cement and the fluidity of concrete. Coarse and fine aggregates contribute similarly (about 25% as influential as curing time) to compressive strength, as they form the primary framework of concrete, despite being bonded by cement and water. Admixtures such as super-plasticizers, slag, and fly ash have the least importance (10%–20% as influential as curing time) since their roles primarily involve improving or
adjusting the performance of concrete or meeting specific design requirements. As discussed, these sensitivities align with engineering practices and the underlying physical principles.
The study examines how the number and sensitivity of input variables affect model performance. Six different combinations of inputs were tested, ranging from the full dataset to just four variables. The results indicate that the full dataset (Combination 1: [cement (kg/m³), water (kg/m³), coarse aggregate (kg/m³), fine aggregate (kg/m³), super plasticizer (kg/m³), blast furnace slag (kg/m³), fly ash (kg/m³), curing time (days)]) produces the best performance. In contrast, the combination with the fewest inputs (Combination 2: [cement (kg/m³), water (kg/m³), coarse aggregate (kg/m³), fine aggregate (kg/m³)]) achieves the lowest performance, with an R² value of 0.377.
Interestingly, simply adding more input variables does not always enhance accuracy. For instance, Combination 3, which includes seven inputs (cement (kg/m³), water (kg/m³), coarse aggregate (kg/m³), fine aggregate (kg/m³), super plasticizer (kg/m³), blast furnace slag (kg/m³), and fly ash (kg/m³)), performs poorly because it omits curing time. Combinations that include key variables, especially curing time and cement, lead to better results. Neglecting curing time significantly undermines accuracy, while the omission of cement has a lesser effect. Ultimately, both curing time and cement are essential for achieving high prediction accuracy.
The AdaBoost model efficiently predicts the concrete compressive strength of various mix proportions over time, helping to verify if designs meet strength requirements and analyze strength growth rates during construction and service phases. It can also serve as a baseline for multi-output models that predict multiple concrete properties simultaneously, such as strength and slump. This approach can optimize mix
proportions to achieve desired strength and slump characteristics, and these applications will be explored further in the future.
Reference
Feng, D. C., Liu, Z. T., Wang, X. D., Chen, Y., Chang, J. Q., Wei, D. F., & Jiang, Z. M. (2020). Machine learning-based compressive strength prediction for concrete: An adaptive boosting approach. Construction and Building Materials, 230, 117000.