Automated gradation design of natural waste gravel soil stabilized by composite soil stabilizer based on a novel DNNSS-APDM-PFC model.

The utilization of natural waste gravel soil as base course material contributes to environmental protection and carbon emission reduction. The purpose of this research is to establish a new model for automated gradation design of the composite soil stabilizer-stabilized waste gravel soil (CSSWGS). A gradation range of CSSWGS has been proposed. The bearing capacity of the waste gravel soils was analyzed using the Particle Flow Code (PFC). The pavement structure performances of CSSWGS with different gradations were also evaluated using the asphalt pavement design method in China (APDM). A critical scientific challenge is to provide foundational predictive data for the gradation design. To address this, a deep learning neural network for small sample (DNNSS) was constructed to predict unconfined compressive strength (UCS) and frost resistance, offering analytical data for both of the aforementioned software. The Adaptive Moment Estimation (Adam) algorithm was employed to dynamically adjust the learning rate, thereby accelerating network; the Dropout function was used to alleviate overfitting; and the Rectified Linear Unit (ReLU) function was used as the activation function to solve the gradient vanishing problem. The results show that the DNNSS algorithm exhibits superior prediction performance compared to other deep learning algorithms. When employing the web version of APDM and the virtual California Bearing Ratio (CBR) test, the analysis results based on the predicted values from DNNSS and measured values were found to be consistent or closely aligned. Consequently, the new DNNSS-APDM-PFC model, leveraging the intelligent algorithm developed in this study, can be effectively utilized for designing the gradations of CSSWGS or analyzing the gradation performances of CSSWGS obtained from field applications.