Probabilistic Machine Learning for Predicting Desiccation Cracks in Clayey Soils
We are thrilled to announce the publication of our recent paper titled “Probabilistic machine learning for predicting desiccation cracks in clayey soils.” In this groundbreaking research, we have developed a probabilistic machine learning (ML) framework to enhance the accuracy and reliability of deterministic models in predicting desiccation cracks in clayey soils, which have become increasingly prevalent due to climate change-induced heatwaves and drought-downpour cycles.
To achieve this, we utilized a comprehensive set of data-driven soil and environment parameters, including initial water content, crack water content, final water content, soil layer thickness, temperature, and relative humidity, as inputs to predict the crack surface ratio. We then employed various ML techniques, such as random forests, regression trees, gradient-boosted trees, support-vector machines, and artificial neural network-particle swarm optimization, to develop predictive models.
To account for uncertainties, we incorporated Monte Carlo simulation, which involved shuffling and randomizing samples, into our models. Additionally, we conducted sensitivity analyses, including input exclusion and partial dependence-individual conditional expectation plots, to assess the reliability of our predictions.
The results of our study revealed that the performance ranking of the developed ML models can be summarized as follows: support-vector machines > gradient-boosted trees > extreme gradient-boosted trees > artificial neural network-particle swarm optimization > random forests > regression trees. However, when considering the probabilistic modeling based on Monte Carlo simulation, gradient-boosted trees demonstrated the highest capability for predictions, exhibiting the lowest errors and uncertainties.
Furthermore, our sensitivity analyses highlighted the importance of various parameters, with the order of significance being final water content > crack water content > soil layer thickness > initial water content > temperature > relative humidity.
A berief review of the research
The importance of the research
The research presented in this paper is of significant importance in the field of geotechnical engineering, particularly in the context of climate change and its impact on soil behavior. The study addresses the issue of desiccation cracks in clayey soils, which have become increasingly prevalent due to heatwaves and drought-downpour cycles associated with climate change.
The paper introduces a novel approach by developing a probabilistic machine learning (ML) framework to enhance the accuracy and reliability of deterministic models used for predicting desiccation cracks. This represents a significant advancement in the field, as traditional deterministic models often struggle to capture the complex and uncertain nature of soil behavior under changing environmental conditions.
The research objectives are highly relevant and aligned with the current challenges faced by geotechnical engineers. By incorporating a comprehensive set of data-driven soil and environment parameters, the study aims to improve predictions of crack surface ratio (CSR), which is a crucial factor in assessing the vulnerability of clayey soils to desiccation cracks.
The use of various ML algorithms, including ensemble regression trees, gradient-boosted trees, support-vector machines, and artificial neural networks, demonstrates the ambition of the research. By exploring multiple ML techniques, the study goes beyond the state of the art and provides a comprehensive comparison of their performance in predicting CSR.
Furthermore, the incorporation of Monte Carlo simulation (MCS) to introduce uncertainties in the models is a significant innovation. This approach allows for the assessment of prediction reliability by quantifying prediction error variances. The sensitivity analyses conducted in the study, such as input exclusion and partial dependence-individual conditional expectation plots, further enhance the understanding of the model’s performance and parameter importance.
The paper introduces a novel approach by developing a probabilistic machine learning (ML) framework to enhance the accuracy and reliability of deterministic models used for predicting desiccation cracks. This represents a significant advancement in the field, as traditional deterministic models often struggle to capture the complex and uncertain nature of soil behavior under changing environmental conditions.
The research objectives are highly relevant and aligned with the current challenges faced by geotechnical engineers. By incorporating a comprehensive set of data-driven soil and environment parameters, the study aims to improve predictions of crack surface ratio (CSR), which is a crucial factor in assessing the vulnerability of clayey soils to desiccation cracks.
The use of various ML algorithms, including ensemble regression trees, gradient-boosted trees, support-vector machines, and artificial neural networks, demonstrates the ambition of the research. By exploring multiple ML techniques, the study goes beyond the state of the art and provides a comprehensive comparison of their performance in predicting CSR.
Furthermore, the incorporation of Monte Carlo simulation (MCS) to introduce uncertainties in the models is a significant innovation. This approach allows for the assessment of prediction reliability by quantifying prediction error variances. The sensitivity analyses conducted in the study, such as input exclusion and partial dependence-individual conditional expectation plots, further enhance the understanding of the model’s performance and parameter importance.
Additional information
Due to space limitations, the complete codes and data are not included in the paper. Nevertheless, upon reasonable request, the authors are willing to provide the interested readers with the necessary sample code and data to facilitate further understanding and replication of the study.
Final words
We invite our readers to contact us if they have any comments or questions regarding this paper. We also encourage researchers to freely cite our paper in their own research, as we believe that collaboration and knowledge sharing are crucial for further advancements in this field.
Thank you for your interest in our research, and we look forward to hearing from you.
Link to access the paper: http://dx.doi.org/10.1007/s10064-023-03366-2
سبد خرید