An Artificial Intelligence Based Data-Driven Method for Forecasting Unconfined Compressive Strength of Cement Stabilized Soil by Deep Mixing Technique

摘要

The cost constraints imposed on construction projects specially ground improvements by the deep mixing technique (DMT) highlight the role of efficient optimization. In this study, a systematic artificial neural network (ANN)-based method to address this practical issue is presented. To achieve this goal, the mutual information method is firstly used to select the most important input parameters for training. Secondly, different ANN architectures are examined to find the network with the smallest error by training and optimizing several multi-layer perceptron networks using a grid search-based technique. The model is then generalized by employing the K-fold and Hold-out cross-validation methods. Comparing K-fold and Hold-out methods showed that K-fold has a longer computation time (approximately 4.5 times longer than the Hold-out method) but leads to higher accuracy (R-2 of 0.98 as compared to R-2 of 0.89 for the Hold-out method). Furthermore, the influence of each input variable on the output is examined by sensitivity analyses emphasizing the importance of the water-cement ratio and the cement content as the input parameters. Lastly, the proposed methodology is validated against a large-scale field DMT experiment. The results showed an acceptable accuracy in predicting the strength of soil after DMT with an R-2 of 0.83.