Nanotechnology led to the development of new materials, potentially enhancing concrete mechanics. The intensity of concrete varies with nanomaterial type and concentration. Previous research has shown that reliable prediction for nanomaterial-reinforced concrete strength is lacking.
Study: Boosting Concrete Strength: Machine Learning Models for Nanosilica-Reinforced Concrete. Image credit: I.Friedrich/Shutterstock
In a recent paper published in the journal Engineering Fracture Mechanics, researchers examined 11 machine learning (ML) algorithms to predict the uniaxial compressive strength (UCS) of nanosilica-reinforced concrete.
Background
Recent literature showcases various additive-enhancing materials such as steel, cement, and concrete. This insight can optimize construction by improving mix designs, structural performance, and durability. Understanding the failure mechanics of these materials is essential for informed design and risk mitigation. Nanotechnology has influenced engineering and enhanced materials such as steel and concrete.
New materials from nanotech advancements improve the mechanical behavior and durability of building materials. Researchers conducted experiments with nanosilica, nanoclay, carbon nanotubes, and nano-aluminum to enhance concrete properties and found that carbon nanotubes boost concrete strength by 21%. Additionally, nanosilica's effect on high-performance concrete varies with specific surface area and water-to-binder ratio, and when added in varying concentrations, it boosts strength and reduces porosity. Nanomaterials such as nano-hematite and nanoaluminum were also found to enhance concrete mechanics. However, discrepancies usually arise due to nanomaterial dispersion in cement.
Advanced methods such as empirical regression, numerical simulation, and ML are used to measure concrete strength. ML techniques such as artificial neural networks (ANNs) and support vector machines (SVMs) efficiently predict concrete strength.
Data collection and ML algorithms for concrete strength estimation
ML encompasses supervised, unsupervised, reinforcement, semi-supervised, and active learning. The current study explores 11 ML algorithms to estimate nanosilica-reinforced concrete's compressive strength. These algorithms include a range of techniques, including ANN, random forests (RF), null space SVR (NuSVR), support vector regression (SVR), Gaussian process regression (GPR), extreme gradient boosting (XGBoost), K-nearest neighbors (KNN), hist-gradient boosting regressor (HGBR), voting regressor (VR), gradient boosting regressor (GBR), decision tree regressor (DTR), and extra tree regressor (ETR). ANN mimics biological neural networks, while ensemble techniques enhance prediction accuracy.
To collect data for this research, 460 cylindrical concrete samples were prepared, each containing varying nanosilica percentages. Following the mixing plan, materials were poured into PVC molds, and after 24 hours, the concrete samples were moved to a curing environment for 28-day strength development. Subsequently, they were cut and polished to prepare for UCS testing. Ensuring the credibility of the database is essential for accurate predictions, which requires trustworthy data and a comprehensive range of parameter values.
Given the high data dimensionality, models are built using a subset of predictor variables. Additionally, feature selection, particularly stepwise techniques, aids in dimensionality reduction. Step Akaike information criterion (StepAIC), a feature selection method, is used to construct optimal models.
Five crucial parameters for UCS were identified: nanosilica percentage in concrete cement (NS), sample length (L), sample diameter (D), P-wave velocity (Vp), and porosity (n). Principal component analysis (PCA) was employed to analyze and streamline dimensionality. The results showed that the top two principal components explained 80% of the dataset's variation, although distinct natural groupings were lacking.
Quantification of ML model performance
Statistical assessment criteria were employed to evaluate model performance. These criteria include the coefficient of determination (R2), mean absolute error (MAE), mean squared error (MSE), mean absolute percentage error (MAPE), and a10-index. Hyperparameters determine model parameter values learned during training. To optimize performance, hyperparameters are fine-tuned iteratively, using trial and error. After fine-tuning the ML models using training data, their performance was evaluated using test data.
Numerical measures provide insights into model performance. Various models can predict UCS with a10-index value above 0.88. Additional statistical metrics such as MAE, MSE, MAPE, and R2 are used to evaluate model performance. Based on accuracy, ETR is the best performer, while DTR is the least performer.
Considering the effect of nanosilica on concrete UCS, models are developed for predicting UCS variations. Some models exhibit consistent behavior, while others show mixed trends. The most suitable models, SVR and NuSVR, predict UCS based on different nanosilica percentages. These models reveal UCS increases up to a certain nanosilica percentage, followed by a decrease. As per the observations, the optimal models for predicting concrete's UCS are SVR and NuSVR. These models indicate how different parameters impact UCS, highlighting nanosilica's role in enhancing concrete strength.
Conclusion
In summary, researchers explored different ML algorithms to predict the UCS of nanosilica-reinforced concrete and concluded that parameters NS, n, Vp, D, and L significantly influence concrete's UCS. The findings showed that SVR and NuSVR models best predict UCS changes. The optimal nanosilica content for maximum UCS is around 20%. NS, n, and Vp highly affect UCS, while D and L have smaller impacts.
Future research should develop models to assess and predict strength characteristics with diverse additives, enhancing the construction industry's adaptability. Researchers should apply the current models to estimate the mechanical properties of different materials, considering factors such as feature selection, data collection, model training, validation, and iterative refinement for accurate predictions of specific mechanical parameters. Numerous studies confirm the effectiveness of ML methods for estimating material properties, highlighting their potential for future investigations into predicting various mechanical properties.
Machine Learning Models for Turbulent Combustion Speed Prediction