This study systematically presents the application of machine learning (ML) algorithms for constructing a constitutive model for soils. A genetic algorithm is integrated with ML algorithms to determine the global optimum model, and the k-fold cross-validation method is used to enhance the models’ robustness. Three typical ML algorithms with formulations explicitly expressed [i.e., back-propagation neural network (BPNN), extreme learning machine (ELM) and evolutionary polynomial regression (EPR)], and two modelling strategies (i.e. total or incremental stress–strain strategies) are used. A synthetic database is first generated based on a simple constitutive model to objectively evaluate the performance of three ML algorithms and two modelling strategies. Next, the optimum ML algorithm and the well evaluated modelling strategy are applied to experimental tests for examining its robustness. All results indicate that a BPNN-based constitutive model using the incremental stress–strain strategy performs best in modelling the mechanical behaviour of soils in terms of interpolation and extrapolation abilities, followed by ELM and then EPR.

本研究系统地介绍了机器学习 (ML) 算法在构建土壤本构模型中的应用。遗传算法与ML算法相结合，确定全局最优模型，k-fold 交叉验证方法用于增强模型的鲁棒性。三种具有明确表达公式的典型 ML 算法 [即，反向传播神经网络 (BPNN)、极限学习机 (ELM) 和进化多项式回归 (EPR)] 和两种建模策略（即总或增量应力-应变策略）是用过的。首先基于简单的本构模型生成合成数据库，以客观评估三种 ML 算法和两种建模策略的性能。接下来，将最优 ML 算法和经过良好评估的建模策略应用于实验测试以检查其鲁棒性。