Prediction of loess collapsibility coefficient using bayesian optimized random forest model.

Area of Science:

• Geotechnical Engineering
• Earth Sciences
• Computational Science

Background:

• Accurate prediction of the collapsibility coefficient of loess is critical for mitigating engineering hazards and understanding environmental impacts.
• Traditional methods for determining the collapsibility coefficient are inefficient, requiring significant time, labor, and resources.
• Machine learning approaches show promise for predicting loess collapsibility, but hyperparameter optimization has been limited.

Purpose of the Study:

• To comprehensively optimize hyperparameters for machine learning models predicting loess collapsibility using Bayesian optimization.
• To evaluate and compare the performance of six different regression models on both training and independent testing datasets.
• To identify a robust and reliable machine learning model for accurate loess collapsibility prediction.

Main Methods:

• Employed Bayesian optimization for fine-tuning hyperparameters of six distinct regression models.
• Evaluated model performance using R² values on both training and independent testing sets.
• Utilized a Random Forest-based model as a primary candidate for prediction.

Main Results:

• The Random Forest-based model demonstrated superior performance, achieving R² values of 0.915 on the training set and 0.965 on the independent testing set.
• Bayesian optimization significantly improved hyperparameter tuning compared to previous studies.
• The developed model shows high reliability in predicting the collapsibility coefficient of loess.

Conclusions:

• The Random Forest model, optimized with Bayesian techniques, provides a reliable and accurate method for predicting loess collapsibility.
• This machine learning approach offers a more efficient alternative to traditional methods for assessing loess collapsibility.
• The findings contribute to improved hazard mitigation strategies in loess-prone engineering projects.