We have extensively performed the application of AI in geotechnics due to the strong capacity of solving non-linear and high-dimensional problem of AI. For example, the optimization and Bayesian-based methods can bridge the gap between advanced constitutive theories and engineering practice; ML-based surrogate model can also be applied in engineering practice as the alternative to experimental and numerical methods, saving expenses for engineering design. Currently, we focuses on the following three topics.
Related PhD students ｜ Hai-Lin WANG
We enhanced several optimization algorithms (DE, GA, BSA…), applied to intelligent properties correlation, the parameters identification and model selection based on laboratory tests or engineering measurements, and applied to excavation project with intelligent updating of the soil model with parameters.
1.1 Optimization theory: Enhanced backtracking search algorithm
We enhanced an enhanced backtracking search algorithm (so-called MBSA-LS) for parameter identification with two modifications: (1) modifying the mutation of original BSA considering the contribution of current best individual for accelerating convergence speed and (2) novelly incorporating an efficient differential evolution (DE) as local search for improving the quality of population.
1.2 Application to properties correlation: intelligent EPR
We proposed a robust and effective evolutionary polynomial regression (EPR) model for mechanical properties of clay using a newly enhanced differential evolution (DE) algorithm with two enhancements: (1) a new fitness function is proposed using the structural risk minimization (SRM) with L2 regularization that penalizes polynomial complexity, and (2) an adaptive process for selecting the combination of involved variables and size of polynomial terms is incorporated.
1.3 Application to parameters identification, model selection and updating
We proposed a novel optimization-based intelligent model selection procedure in which parameter identification is also performed during staged excavation. As the process of optimization goes on, the soil model exhibiting good performance during simulation survives from the database and model parameters are also optimized. For each excavation stage, with the selected model and optimized parameters, wall deflection and ground surface settlement of the subsequent unexcavated stage are predicted.
1.4 Optimization tool for parameters identification
In order to conveniently solve the problems of parameter identification in geotechnical engineering, we proposed ErosOpt, that is a practically useful tool. ErosOpt makes it easy to specify and solve geotechnical optimization problems without expert knowledge, and at the same time it allows experts to implement advanced algorithmic techniques or advanced constitutive models.
Related PhD students ｜ Hai-Lin WANG
We proposed the Differential Evolution Transitional MCMC (DE-TMCMC) method. To realize the intended computational savings, a parallel computing implementation of DE-TMCMC was developed using the Single Program/Multiple Data (SPMD) technique in MATLAB, and applied the Parallel-based DE-TMCMC for parameter identification and model class selection.
Example 1: Parameter identification from PMT by DE-TMCMC
Example 2: Model class selection by DE-TMCMC
Related PhD students ｜ Geng-Fu HE ｜ Hao-Ran ZHANG ｜ Hao-Han HUANG
3.1 ML-based correlations of soil properties
We adopted a series of machine learning algorithms (e.g. GP = genetic programming; EPR = evolutionary polynominal regression; SVR = support vector regression; RF = random forest; BPNN = backpropagation neural network; ANN_MCD = artificial neural network and Monte Carlo dropout) to propose prediction models for soil properties (e.g. compressibility index of clay, creep index of clay, undrained shear strength of clay, hydraulic conductivity of clay, soil compaction parameters, air-entry value of compacted soils). A machine learning based GUI was also developed for easy to use.
3.2 ML-based image reconstruction and identification of geomaterials
A hybrid machine learning algorithm was proposed to segment computed tomography (CT) images then kernel-based algorithm combined with level set was proposed to automatically reconstruct particles in a CT image. The reconstructed particle was fed to a three-dimensional convolutional neural network based particle shape prediction model to obtain commonly used shape parameters such as aspect ratio, roundness, circularity and convexity.
3.3 ML-based modelling of soil behaviours
We adopted a series of machine learning algorithms to model path-dependent soil behaviour and found that the long short-term memory neural network outperforms other machine learning algorithm in this problem. LSTM-based data-driven model has been successfully applied to model monotonic and cyclic soil behaviour on Baskarp sand, Fontainebleau sand and Toyoura sand, as well as from a soil sample image to directly extract macro-scale mechanical behaviour.
3.4 Application of ML models in engineering practice
Data-driven based constitutive model was integrated with finite element method for modelling engineering-scale problem. A physics-informed learning mechanism was invoked to improve the generalization ability of the data-driven model. To reduce the dependency on the high-fidelity data and architecture of the neural network, a multi-fidelity residual training framework was proposed and used to develop surrogate model for engineering practice.
