基于余能原理基面力元法与极限学习机相结合的轻骨料混凝土抗拉强度预测

Prediction of Tensile Strength of Lightweight Aggregate Concrete Using a Baseline Element Method Based on Residual Energy Principle Combined with a Limit Learning Machine

摘要: 轻骨料混凝的抗拉强度是其关键力学性能，但受骨料、界面过渡区、孔隙等多相细观结构影响显著，传统经验公式或宏观数值模拟预测精度有限。本文提出一种细观与宏观混合的预测模型。首先，利用基面力元法这一高效处理不连续问题的数值方法，构建轻骨料混凝土的二维细观模型。通过引入随机骨料投放算法真实模拟了拉伸作用下骨料与基体界面开裂与裂纹扩展的全过程。采用正交试验设计，系统研究了试件尺寸、骨料强度、界面过渡区强度、孔隙率等九个关键细观参数对宏观抗拉强度的定量影响，生成了包含130组样本的数据集。随后，将该数据集随机划分为训练集(80组)、验证集(25组)和测试集(25组)，输入到极限学习机神经网络中进行训练与优化。研究结果表明：所提出的BFEM-ELM混合模型能充分挖掘细观结构特征与宏观性能间的复杂非线性关系，预测速度快、精度高，为轻骨料混凝土的材料设计与性能优化提供了新思路。

Abstract: The tensile strength of lightweight aggregate concrete is a key mechanical property, yet it is significantly influenced by the multiphase microstructure involving aggregates, the interfacial transition zone, and pores. Traditional empirical formulas or macroscopic numerical simulations exhibit limited predictive accuracy. This paper proposes a hybrid micro-macro prediction model. First, a two-dimensional micro-scale model of lightweight aggregate concrete is constructed using the base-surface force element method, a highly efficient numerical approach for handling discontinuities. By incorporating a random aggregate placement algorithm, the model realistically simulates the entire process of interface cracking and crack propagation between aggregates and matrix under tensile loading. Employing an orthogonal experimental design, the study systematically investigates the quantitative influence of nine key micro-parameters—including specimen size, aggregate strength, interfacial transition zone strength, and porosity—on macro-tensile strength, generating a dataset comprising 130 sample sets. Subsequently, this dataset was randomly partitioned into training (80 sets), validation (25 sets), and testing (25 sets) sets, then input into an Extreme Learning Machine (ELM) neural network for training and optimization. Results demonstrate that the proposed BFEM-ELM hybrid model effectively uncovers the complex nonlinear relationship between microstructural features and macroscopic performance, delivering fast prediction speeds and high accuracy. This approach offers novel insights for material design and performance optimization in lightweight aggregate concrete.