基于机器学习算法的非饱和土水特征曲线预测

Prediction of Unsaturated Soil Water Characteristic Curve Based on Machine Learning Algorithms

摘要:

土水特征曲线（SWCC）是研究非饱和土渗透、强度预测与本构关系的基础。机器学习算法具有高效处理大量数据和特征提取等特点。采用六种机器学习算法（四种集成学习和两种传统机器学习算法）对美国非饱和土数据库中的154条SWCC包含1976个数据点进行模拟；并使用四个性能评价指标（R²、EVS、MAE和RMSE）评价算法的性能。选取两种数据输入的方式：对压力水头进行对数处理和未处理两类。结果表明，在两种输入情况下，对 LightGBM、XGB、RF和AdaBoost算法的影响很小；但是对GPR和SVM两种传统机器学习算法的影响很大，在未进行对数处理情况下，R²降低明显甚至会出现无法模拟SWCC的情况。此外，LightGBM对SWCC测试集的模拟效果上均优于其他模型，拥有高的趋势评价指标（R²和EVS）和低的误差测量指标（MAE和RMSE）；六种算法对 SWCC模拟的优劣的排列顺序依次为：LightGBM、GPR、XGB、RF、AdaBoost和SVM。最后，利用已训练好的 LightGBM模型对9条不包含在数据库内的SWCC数据进行预测，结果显示LightGBM能够较好地预测非饱和土的土水特性。研究结果对提升不同类型土的SWCC预测具有重要的指导意义。

Abstract:

The soil water characteristic curve (SWCC) is fundamental for studying the permeability, strength prediction, and constitutive relationships of unsaturated soils. Machine learning algorithms are characterized by their efficiency in large dataset processing and feature extraction. This study used six machine learning algorithms (four ensemble learning and two traditional machine learning algo rithms) to simulate 154 SWCCs with 1976 data points from the United States Unsaturated Soil Data base. Four performance evaluation indicators (R², EVS, MAE, and RMSE) were used to assess the algorithms' performance. Two types of data input methods were selected: one with logarithmic pro cessing of matric suction, and the other without any transformation. The results showed that, under both input types, the effect on the LightGBM, XGB, RF, and AdaBoost algorithms was minimal. However, the two traditional machine learning algorithms, GPR and SVM, were significantly affect ed. Without logarithmic transformation, R² decreased noticeably, and in some cases, the SWCC could not be simulated. Additionally, LightGBM outperformed other models in simulating the SWCC for the test set, with higher trend evaluation indicators (R² and EVS) and lower error measurement in dicators (MAE and RMSE). The ranking of the six algorithms in terms of SWCC simulation perfor mance was as follows: LightGBM, GPR, XGB, RF, AdaBoost, and SVM. Finally, the trained LightGBM model was used to predict 9 SWCC datasets not included in the original database. The re sults showed that LightGBM could effectively predict the soil water characteristics of unsaturated soils. The findings provide important guidance for improving SWCC predictions for different types of soils.