Machine Learning Classification for Acoustic Mix Identification in Styrofoam Lightweight Concrete: A Feature Sensitivity Study

Abstract

Concrete has long been the backbone of modern construction, yet the demands placed on building materials have changed considerably. Strength and structural integrity remain essential, but acoustic performance, how effectively a structure manages sound, has become a serious design consideration. This study takes that challenge as its starting point, examining Styrofoam-based lightweight concrete in which expanded polystyrene beads partially replace fine aggregate to improve sound absorption. Four concrete mixtures (FS, ST1, ST2, and ST3), were prepared with progressively increasing Styrofoam content and characterised for compressive strength, flexural strength, and sound absorption coefficient (α) using the impedance tube method (ASTM E1050-12). A dataset of 312 records was compiled from these four unique material compositions, each tested across 13 frequency levels (100–1600 Hz) and six acoustic panel configurations.
Rather than predicting continuous acoustic absorption values, this study focuses on identifying between concrete mix types based on their measured material and acoustic characteristics using machine learning (ML) classification models. The acoustic behaviour of Styrofoam-based lightweight concrete is shaped by the nonlinear interaction of pore geometry, bead size distribution, and frequency-dependent viscous losses relationships that conventional empirical models cannot easily capture. ML offers a data-driven route to material identification that sidesteps those modelling constraints. Four algorithms k-Nearest Neighbours (k-NN), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Logistic Regression (LR), were evaluated using 10-fold cross-validation in Orange Data Mining 3.40.0, through two complementary analyses: Analysis 1 with the full feature set, and Analysis 2 restricted to acoustic features only.
In Analysis 1, ANN achieves perfect CA = 1.000 and AUC = 1.000, while SVM reaches CA = 0.923. However, both k-NN and Logistic Regression return CA = 0.250, identical to random chance, due to the absence of internal feature normalisation in Orange. A feature sensitivity analysis confirms that ANN's perfect result is mechanistically explained by class-unique compositional features, not acoustic learning. In Analysis 2, removing those identifying features drops all models to or near random chance (CA = 0.21–0.25; balanced baseline = 0.25). A regression baseline confirms that continuous α prediction is not currently feasible (R² ≤ −0.32). The main contribution lies not in the accuracy numbers but in the dual-analysis evaluation framework, a transparent approach that clearly separates what ML can and cannot do with small experimental datasets. Experimentally, ST2 (1:1.5:2.5) emerged as the most balanced mix.