LSTM-based deep learning model for alkali activated binder mix design of clay soils

Alkali-activated binders have been extensively studied for ground improvement, particularly in cohesive soils. However, the adoption of this method in practice is often hindered by the complex mix design process, owing to the multitude of factors influencing the compressive strength of the treated soil. The mechanical performance of the AAB-treated soil mix is not solely governed by the binder properties but is also significantly affected by the inherent soil properties, which can vary widely across different locations. This study explores the development and optimization of deep learning (DL) models for the design of alkali-activated binder (AAB) treated clay mixtures. Leveraging a large dataset of experimental test records, several DL techniques, including Linear Regression, Support Vector Regression, Artificial Neural Network, and Long Short-Term Memory (LSTM), were utilized to predict the unconfined compressive strength of AAB-treated clay based on various input parameters. Among all the models evaluated, LSTM provided the highest predictive accuracy with a determination coefficient, root means square error, and mean absolute error of 75.8%, 0.696, and 0.412, respectively. Sensitivity analysis revealed soil plasticity and alkali activator-to-binder ratio as the most significant parameters affecting AAB-treated clay compressive strength. The study concludes that the LSTM model developed can serve as a practical tool for designing AAB-treated clay mixes and understanding the key factors influencing their compressive strength.