Enhancing unconfined compressive strength prediction in nano-silica stabilized soil: a comparative analysis of ensemble and deep learning models

The study emphasizes the challenges of determining the Unconfined Compressive Strength (UCS) of soil stabilized using nano-silica (NS) in civil engineering applications. As a result, a thorough strategy combining three ensemble learning (EL) and deep learning (DL) algorithms was created, and it was discovered that the best DL model was Long Short-Term Memory (LSTM) and the most accurate EL model was Gradient Boosting (GBR). With R2 values of 1.0 for training and 0.9684 for testing datasets, along with a low Root Mean Square Error (RMSE) of 0.0203, the GBR model demonstrated remarkable accuracy. Similar to this, LSTM models demonstrated remarkable accuracy, with RMSE values of 0.022 and R2 values of 0.9819 and 0.9405 for training and testing datasets, respectively. The models' practical utility in geotechnical engineering was confirmed by the Bland-Altman analysis, which revealed minor mean differences for both models. Furthermore, the GBR model was computationally more efficient than the LSTM. The effectiveness of the models was further shown by validation against a sizable number of UCS experiment trials, yielding R2 values of 0.94 and 0.93 for GBR and LSTM, respectively. These results highlight the accuracy, flexibility, and resilience of the GBR model, providing substantial time and cost savings for accurate UCS prediction in NS-stabilized soil and enabling civil engineering professionals to design and build infrastructure with optimal efficiency.