Scholarlink

<p>The development of reliable predictive models for geopolymer concrete compressive strength remains a critical challenge in sustainable construction materials. Although machine learning provides promising solutions, conventional models such as the Multilayer Perceptron Neural Network (MLPNN) and Gaussian Process Regression (GPR) suffer from limitations related to parameter optimization and computational efficiency. This study proposes a systematic hybrid modelling framework that synergistically integrates three nature-inspired metaheuristic algorithms, namely the Mayfly Optimization Algorithm (MOA), Artificial Protozoa Optimizer (APO), and Stellar Oscillation Optimizer (SOO), with base ML models to develop six advanced hybrid predictors: MOA-MLPNN, APO-MLPNN, SOO-MLPNN, MOA-GPR, APO-GPR, and SOO-GPR. Particle Swarm Optimization (PSO) was additionally incorporated as a well-established baseline optimizer to enable rigorous comparative benchmarking. Comprehensive evaluation shows that all metaheuristic-augmented models substantially outperform their standalone counterparts, confirming the effectiveness of optimization-driven learning in capturing the highly nonlinear behavior of geopolymer concrete. Among the evaluated models, the MOA-optimized MLPNN demonstrated the most reliable performance, achieving the highest goodness-of-fit indices (R² = 0.95914, KGE = 0.9370, NSE = 0.9586) and the lowest prediction errors (MSE = 0.00358, MAPE = 2.15%). The results further indicate that advanced optimizers provide measurable advantages over classical approaches such as PSO when strategically paired with suitable base learners. To enhance interpretability, SHAP-based analysis was employed, revealing that silica fume and natural zeolite contents are the dominant factors governing compressive strength prediction, while curing time and sodium hydroxide (NaOH) molarity play secondary roles within the studied ranges. The findings demonstrate that strategic optimizer–model pairing, rather than hybridization alone, is the key driver of performance enhancement. The MOA-MLPNN hybrid therefore offers a robust, interpretable, and high-precision tool for geopolymer mix design optimization and sustainable construction practice.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid metaheuristic-machine learning framework for high-fidelity prediction of geopolymer concrete compressive strength

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Godwin Ativor,
  • Samuel Nana Eshun,
  • Prince Kofi Otchere,
  • Yao Yevenyo Ziggah

摘要

The development of reliable predictive models for geopolymer concrete compressive strength remains a critical challenge in sustainable construction materials. Although machine learning provides promising solutions, conventional models such as the Multilayer Perceptron Neural Network (MLPNN) and Gaussian Process Regression (GPR) suffer from limitations related to parameter optimization and computational efficiency. This study proposes a systematic hybrid modelling framework that synergistically integrates three nature-inspired metaheuristic algorithms, namely the Mayfly Optimization Algorithm (MOA), Artificial Protozoa Optimizer (APO), and Stellar Oscillation Optimizer (SOO), with base ML models to develop six advanced hybrid predictors: MOA-MLPNN, APO-MLPNN, SOO-MLPNN, MOA-GPR, APO-GPR, and SOO-GPR. Particle Swarm Optimization (PSO) was additionally incorporated as a well-established baseline optimizer to enable rigorous comparative benchmarking. Comprehensive evaluation shows that all metaheuristic-augmented models substantially outperform their standalone counterparts, confirming the effectiveness of optimization-driven learning in capturing the highly nonlinear behavior of geopolymer concrete. Among the evaluated models, the MOA-optimized MLPNN demonstrated the most reliable performance, achieving the highest goodness-of-fit indices (R² = 0.95914, KGE = 0.9370, NSE = 0.9586) and the lowest prediction errors (MSE = 0.00358, MAPE = 2.15%). The results further indicate that advanced optimizers provide measurable advantages over classical approaches such as PSO when strategically paired with suitable base learners. To enhance interpretability, SHAP-based analysis was employed, revealing that silica fume and natural zeolite contents are the dominant factors governing compressive strength prediction, while curing time and sodium hydroxide (NaOH) molarity play secondary roles within the studied ranges. The findings demonstrate that strategic optimizer–model pairing, rather than hybridization alone, is the key driver of performance enhancement. The MOA-MLPNN hybrid therefore offers a robust, interpretable, and high-precision tool for geopolymer mix design optimization and sustainable construction practice.