Zhang Wan, Guo Jiangtao, Li Zhaopeng, Cheng Ruifang, Ning Cuiping, Niu Hongfeng, Liu Ze
College of Architecture Engineering, Yangling Vocational & Technical College, Yangling, 712100, Shaanxi, China.
The First Affiliated Hospital of Xinxiang Medical University, Xinxiang, 453003, Henan, China.
Sci Rep. 2025 Jul 12;15(1):25281. doi: 10.1038/s41598-025-11121-8.
Accurately predicting the collapsibility coefficient of loess is crucial for mitigating the hazards associated with loess collapsibility in engineering projects, natural environment, and socio-economic activities. The traditional method for determining the collapsibility coefficient is time-consuming, labor-intensive, and expensive. In recent years, researchers have increasingly employed machine learning techniques to predict collapsibility coefficient and have obtained promising results. However, the process of hyperparameter optimization in previous studies was not sufficiently comprehensive, leading to suboptimal model performance. Therefore, in this study, Bayesian optimization was employed to fine-tune the hyperparameters of six different regressors, and the performance of these models was evaluated on both a training set and an independent testing set. The results demonstrated that the Random Forest-based model achieved the best performance, with R² values of 0.915 and 0.965 on the training and independent testing sets, respectively. These findings indicate that the proposed model is capable of reliably predicting the collapsibility coefficient of loess.
准确预测黄土的湿陷系数对于减轻工程项目、自然环境和社会经济活动中与黄土湿陷性相关的危害至关重要。传统的确定湿陷系数的方法既耗时、费力又昂贵。近年来，研究人员越来越多地采用机器学习技术来预测湿陷系数，并取得了有前景的结果。然而，以往研究中的超参数优化过程不够全面，导致模型性能欠佳。因此，在本研究中，采用贝叶斯优化对六种不同回归器的超参数进行微调，并在训练集和独立测试集上评估这些模型的性能。结果表明，基于随机森林的模型表现最佳，在训练集和独立测试集上的R²值分别为0.915和0.965。这些发现表明，所提出的模型能够可靠地预测黄土的湿陷系数。
