In order to more accurately describe the behavior of geological materials, constitutive models with increasing complexity are being developed. The increase in the number of parameters and equations makes engineering applications difficult. On the other hand, data-driven machine learning approaches have shown great potential in addressing this issue. Therefore, from a data perspective, this paper recapitulates and discuss the task and framework of machine learning. For the parameter calibration task, 1-Dimensional Convolutional Neural Networks (1D-CNN), Recurrent Neural Network (RNN), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO) were tested under the framework of representation learning and optimization. In the preset computing environment, the normalized RMSEs of representation learning are 0.21 and 0.17, and the calculation takes less than 0.1 second. Optimization frameworks perform better with RMSEs of 0.04 and 0.07, but cost over 10 hours. Finally, the advantages and disadvantages of each framework are discussed.
The mechanical behavior of geomaterials is often observed with nonlinear laws which have been shown to involve structural properties, anisotropy, saturation, temperature, chemistry, and time dependence. These complex stress-strain relationships, which are usually systematically and controllably generalized by laboratory element experiments, are mathematically described as constitutive model. Using constitutive model under geometric and mechanical boundary conditions, it is feasible that establish numerical simulations for solving engineering problem. In order to provide high quality computational simulations and recommendations in specific geological conditions, researchers have been exploring and creating new constitutive models to more accurately describe geological materials. In the past few decades, from the original perfect linear-elastic model, it was developed to include more accurate critical state models for complex geological states. However, there are still several issues that limit the development and application of constitutive models: 1) As the mathematical formulation becomes complex, the parameter calibration of the constitutive model is very cumbersome causing difficulties in application. 2) Many models are proprietary only for specific geological conditions and show low applicability to other geological conditions.
