Generalized uncertainty in surrogate models for concrete strength prediction

In this study, a new method has been developed using Convolutional Neural Networks (CNN) and Support Vector Regression (SVR) integration to determine the compressive strength of concrete, which is the most common structural material used in the construction sector. The developed CNN model (MPaConvNet) extracts important features from concrete surface images. These extracted features are then used to predict the compressive strength of concrete using the SVR algorithm. The dataset used in the study consists of 1453 concrete surface images obtained from 203 core samples collected from 115 real buildings. The proposed method predicted the compressive strength of concrete with the values of 99.59%, 0.0209, 0.6836, 0.4673, and 1.0000 according to the R², MAPE, RMSE, MSE, and a-20 metrics, respectively. Additionally, in the study, the regions focused on by the proposed CNN architecture in concrete surface images were detected using the Gradient-weighted Class Activation Mapping algorithm (Grad-CAM).

High temperature severely deteriorates the durability and mechanical properties of concrete. Experimental tests to obtain the compressive strength of fire-damaged concretes are costly and time-consuming. Soft computing techniques can be used to estimate the residual compressive strength of concretes rapidly and accurately. This study developed three general, optimal and accurate 2D convolutional neural network (CNN) models to predict the compressive strength of fire-damaged concretes. The data set contained 1062 samples, split randomly into the training (60%), validation (20%) and verification (20%) sets. The inputs included mixture proportion, chemical compositions of cementitious materials, specimen shape and size, heating regime (peak temperature, heating rate and exposure time) and age. The residual compressive strength was the output. The 28-day and unheated compressive strengths were the supplementary inputs in two of the models. The genetic algorithm optimized the models' architecture, including the layers’ number and their hyperparameters. All three models showed acceptable accuracy; however, considering the supplementary inputs enhanced the accuracy slightly. The R², RMSE, MAE, MAPE and VAF values of the verification set in the model considered the 28-day compressive strength as the supplementary input feature were 0.965, 4.80 MPa, 3.48 MPa, 12.83%, and 96.35%, respectively. The sensitivity analysis revealed that the heating regime, especially peak temperature, is the most fundamental input feature. Finally, the superiority of the proposed models over the previous studies was shown based on data set size and statistical metrics.

Building precise machine learning and deep learning models has traditionally required a combination of mathematical skills and hands-on experience to meticulously adjust hyperparameters that significantly impact the learning process. As datasets continue to expand across various engineering domains, researchers increasingly turn to machine learning methods to uncover hidden insights that may elude classic regression techniques. This surge in adoption raises concerns about the adequacy of resultant meta-models and the interpretation of findings. In response to these challenges, automated machine learning (AutoML) emerges as a promising solution, aiming to construct machine learning models with minimal intervention or guidance from human experts. This paper benchmarks AutoML solutions by providing an overview of their principles and applying them to predict the most important mechanical properties of different concrete datasets, i.e., compressive strength. Nine datasets from various concrete types, sample sizes, and features are utilized, with a detailed discussion on the benchmark dataset from high-performance concrete, applying best practices to the other eight datasets. For each case, the importance of hyperparameter tuning is discussed, alongside the ensemble and stacking models. Tree-based models are employed for each dataset to develop SHAP plots, interpret results, and understand the contribution of each component in the mix design to the overall strength of the concrete. This paper further explores three unique aspects of benchmarking AutoML in material science: (1) “reliability” by contrasting the benchmark dataset’s error metric with literature collected over the past 20 years, (2) “uncertainty” by quantifying the variability in the mean and standard deviation of the error metric from different datasets and its correlation with the sample-to-feature ratio, and (3) “dilemma” by discussing the shortcomings of AutoML in specific concrete datasets.

This study presents an innovative approach to predict concrete compressive strength using particle packing theories through machine learning techniques. The existing challenge in concrete engineering lies in the accurate estimation of concrete strength, a critical factor in construction. The adoption of particle packing theories, which hold great promise for enhancing concrete performance, has been limited due to the complexity and time-consuming nature of the required calculations. An approach encompassing particle packing models (JD Dewar Model, Compressible Packing Model, and Modified Toufar Model) with machine learning is the novelty of the work. These models optimize the packing density of aggregate proportions while minimizing the void ratio, essential for achieving desired compressive strength criteria. To train the model, a comprehensive dataset comprising 479 concrete mixtures, each associated with known compressive strength values relative to packing density, is utilized. A significant advancement in predicting concrete compressive strength is demonstrated by the results. The approach outperforms traditional empirical models, offering precise and reliable predictions based on packing density. Importantly, this innovation eliminates the need for time-consuming and costly trial-and-error procedures in concrete mix design. The strong performance of various models in predicting concrete strength using particle packing theories is underscored by the study, with R^2 values ranging from 0.664 to 0.999. By combining concepts of particle packing theories and machine learning, a more efficient and reliable method for predicting concrete compressive strength is achieved. This innovation has the potential to revolutionize concrete mix design, leading to more durable and cost-effective construction practices.

This study addresses a critical gap in concrete strength prediction by conducting a comparative analysis of three deep learning algorithms: convolutional neural networks (CNNs), gated recurrent units (GRUs), and long short-term memory (LSTM) networks. Unlike previous studies that employed various machine learning algorithms on diverse concrete types, our study focuses on mixed-design concrete and fine-tuned deep learning algorithms. The objective is to identify the optimal deep learning (DL) algorithm for predicting concrete uniaxial compressive strength, a crucial parameter in construction and structural engineering. The dataset comprises experimental records for mixed-design concrete, and models were developed and optimized for predictive accuracy. The results show that the CNN model consistently outperformed GRU and LSTM. Hyperparameter tuning and regularization techniques further improved model performance. This research offers practical solutions for material property prediction in the construction industry, potentially reducing resource burdens and enhancing efficiency and construction quality.