基于机器学习预测微生物加固钙质砂统一动强度

为探究不同微生物加固程度、有效围压和相对密实度与微生物加固钙质砂的动强度系数之间的相对重要性大小，并建立相应的统一动强度准则，用以预测微生物加固钙质砂的动强度，依据86组微生物诱导碳酸钙沉淀(MICP)技术加固钙质砂的不排水三轴试验数据建立数据集，利用反向传播神经网络(BPNN)和多元自适应回归样条(MARS)方法对数据集进行训练和预测。结果表明：微生物加固程度对动强度系数的影响程度最大，有效围压的影响稍逊之，相对密实度的影响相对较小，其相对重要性大小均值分别是100%、94%与68%,基于机器学习方法对微生物加固钙质砂的动强度系数进行预测的效果优于传统方法建立的统一动强度准则所获得的结果，其中BPNN模型的预测结果更准确，MARS模型的计算效率更高。

Machine-learning-based Prediction of Unified Dynamic Strength of Microbially Induced Calcium Carbonate Precipitation-treated Calcareous Sand

In this study, the relative influence of the biocementation level, effective confining pressure, and relative density on the dynamic strength coefficient of microbially induced calcium carbonate precipitation (MICP) -treated calcareous sand and the corresponding unified dynamic strength criterion for predicting the dynamic strength of the treated sand were investigated. Data obtained from 86 undrained triaxial tests were analyzed using the backpropagation neural network (BPNN) and multivariate adaptive regression splines (MARS) as the training and prediction methods. The results show that the biocementation levels significantly influence the dynamic strength coefficient more than the effective confining pressure and relative density. Their corresponding mean values of relative importance are 100%, 94%, and 68%, respectively. The prediction ability of the machine learning method for the dynamic strength coefficient of MICP-treated calcareous sand is better than that of the traditional unified dynamic strength criterion. The BPNN model yields better optimal prediction results than the MARS model. However, the MARS model has a higher calculation efficiency than the BPNN model.