A new study introduces a hybrid fuzzy neural network model that significantly enhances the accuracy of predicting hardness properties in high-performance concrete by combining fuzzy logic, advanced machine learning, and chaos-based optimization.
Study: Evaluation of high performance concrete hardness properties using fuzzy logic based modeling. Image Credit: OlegRi/Shutterstock.com
Why Hardness Matters in HPC
High-performance concrete (HPC) is known for its enhanced mechanical properties, which allow for greater compressive strength, durability, ductility, and energy absorption compared to conventional mixes. These qualities make it particularly valuable in seismic construction and infrastructure projects where strength and longevity are essential. HPC also allows for more refined, slender structural components, enabling both performance and design flexibility.
To make full use of these benefits, engineers need reliable ways to evaluate properties like slump and compressive strength. These indicators influence not only the material's behavior during construction but also its long-term structural performance. However, accurately predicting them has remained a challenge due to the material’s complex composition.
Modeling Challenges in HPC
Despite years of research at both the micro and macro scale, HPC hasn’t been widely adopted in practice. Its high cost and intricate mix design, often involving multiple interacting components, make standard predictive tools less effective.
Conventional regression models and standalone machine learning (ML) methods have been used to estimate HPC properties, but these approaches often fail to capture the material’s non-linear relationships and are sensitive to data variability. Uncertainty in critical parameters such as water-to-binder ratio, silica fume, and superplasticizer content further complicates predictions.
While ensemble and hybrid ML techniques show promise, few incorporate fuzzy logic for managing uncertainty. Even fewer use metaheuristic optimization strategies like chaos theory to fine-tune model parameters. These gaps limit the practical usefulness and reliability of current HPC prediction models.
A Hybrid Approach: Introducing HFANN
To address these limitations, researchers developed a hybrid fuzzy artificial neural network (HFANN). This model integrates fuzzy logic with three machine learning techniques—multilayer perceptron (MLP), gradient boosting machine (GBM), and support vector regression (SVR). Each algorithm is fine-tuned using chaos game optimization (CGO), a technique inspired by fractal geometry.
The goal was to model HPC’s complex, uncertain properties more effectively, improving prediction accuracy for both slump and compressive strength.
Fuzzy logic was used to preprocess the input data, employing a Sugeno-type rule base and Gaussian membership functions to generate fuzzy features. This helped the model account for uncertainty and variability across samples. Each sample was enriched with 18 fuzzy features, allowing the model to better recognize patterns in the data.
CGO then optimized the hyperparameters of each ML model, adjusting learning rates, layer depth, and other parameters, to minimize prediction error. The final ensemble model combines the strengths of GBM’s error correction, MLP’s capacity for recognizing non-linear trends, and SVR’s precision in regression tasks.
Dataset and Experimental Setup
The researchers tested the HFANN on a dataset of 191 HPC mix designs sourced from prior studies. These mixtures used crushed granite aggregates, Type I Portland cement, a naphthalene-based superplasticizer, and fine silica sand. The slump test was conducted following ASTM C143, while compressive strength was measured after 182 days according to ASTM C39.
To analyze model performance, the study used standard metrics such as root mean squared error (RMSE), coefficient of determination (R2), mean absolute error (MAE), and others, including the A20 index, variance account factor (VAF), and an objective detection metric (OBJ). Statistical analysis, including box plots and heatmaps, was used to explore data distribution and variable relationships.
Results and Insights
The HFANN framework delivered substantial performance improvements. It reduced RMSE for compressive strength predictions by 25 %, achieving an R2 of 0.98 for compressive strength and between 0.97 and 0.99 for slump. These results outperformed conventional models, including an 8 % R2 improvement over radial basis function networks and 5 % over linear regression.
To validate its effectiveness, the study used multivariate analysis of variance (MANOVA) and Tukey’s honestly significant difference (HSD) test. Both confirmed the HFANN’s predictive strength.
Further analysis revealed complex relationships within the data. While some input features had strong linear correlations with target outputs, many showed weak or even opposing patterns. Correlation among input features also pointed to multicollinearity, an issue that often destabilizes simpler models. These findings highlight the importance of a multilayered approach capable of capturing nonlinear dependencies and handling uncertainty effectively.
HFANN addresses these challenges through its architecture: fuzzy preprocessing manages ambiguity in input data; diverse learners uncover hidden patterns; and chaos-based optimization ensures robust parameter tuning. This combination allows the model to deliver accurate predictions even when standard correlations are weak or inconsistent.
Looking Ahead
What makes this study stand out isn’t just the improved accuracy; it’s the way it handles the real-world complexity of working with high-performance concrete. Rather than relying on simplified models or ignoring uncertainty, the HFANN approach builds those challenges into the solution.
For engineers and researchers working on modern infrastructure, this kind of flexible, data-aware modeling could be a practical step toward more reliable and efficient material design.
Journal Reference
Li, Q. (2025). Evaluation of high performance concrete hardness properties using fuzzy logic based modeling. Scientific Reports, 15(1), 1-21. DOI: 10.1038/s41598-025-18283-5, https://www.nature.com/articles/s41598-025-18283-5
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.