Prediction of thermal properties of lightweight concrete from mechanical parameters using GMDH

Abstract Growing concerns about the energy consumption of heating and cooling systems in buildings have led to an increased demand for construction materials with enhanced insulation performance. Accurate determination of the thermal properties of lightweight concrete (LCW) is vital for evaluating its energy performance; however, conventional high-precision tests are both time-intensive and expensive. In this study, the Group Method of Data Handling (GMDH) model is proposed as a tool for directly predicting the thermal properties of LCW from its mechanical properties. For the experimental study, LCWs were produced using waste rubber, pumice, and expanded perlite aggregates as substitutes. Three prediction scenarios were established to accurately estimate thermal conductivity (λ), specific heat (c p ), and thermal diffusivity (α) from key mechanical properties, including compressive strength (σ c ), tensile strength (σ t ), porosity (ϕ), bulk density (ρ), and ultrasonic pulse velocity (upv). Using the GMDH model, mathematical relationships were established between mechanical inputs and thermal outputs, with the most influential parameters identified through layer-by-layer transformations. A topology consisting of two hidden layers, each with three neurons, was adopted, and the model was benchmarked against five machine learning algorithms. The results showed that GMDH provided high predictive accuracy, particularly for thermal conductivity, with an R 2 of 0.9957 and low error metrics. These findings indicate that GMDH can serve as an effective and cost-efficient approach for predicting the thermal behavior of LCW, while supporting the sustainable use of alternative aggregates in energy-efficient construction.