Experimental Study and Machine Learning Aided Modelling of the Mechanical Behaviour of Rammed Earth

Rammed earth is a sustainable building technique for constructing foundations, floors, and walls using natural raw materials such as earth, chalk, lime, with stabilizers like cements. As the proportion of various materials changes, the mechanical properties of rammed earth materials are also varying correspondingly. A series of experimental studies are first conducted to evaluate the effects of different proportions of raw materials including clay, sand, cement, and water under various loading rates on the strength/deformation properties (peak strength, q_f; residual strength, q_res; initial modulus, E_max; secant modulus at 50% peak strength, E₅₀) and stress–strain relationships (σ₁∼ ε₁) of rammed earth. A soft computing method (extreme gradient boosting machine, XGBoost) is then developed to model peak strength, residual strength, initial modulus, secant modulus and entire stress–strain relationships obtained from the experimental studies. Three performance metrics including the root mean squared error, variance accounted for and R-squared value (R²) are used to measure the performance of the applied model. Comparisons between simulations and experiments show that the developed XGBoost algorithm is a promising alternative in modelling key mechanical properties and entire stress–strain relationships for rammed earth. For stress–strain relationships calculated R-squared value for the training set is 0.978 and that for the testing dataset is 0.908. The key factor that most significantly affects the peak strength, residual strength, initial modulus, secant modulus and entire stress–strain relationships for rammed earth can be identified by using the developed soft computing method.