This exploration reviews machine learning (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$ML$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">ML</mi> </mrow> </math></EquationSource> </InlineEquation>) tactics to project the collapse potential (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$CP$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">CP</mi> </mrow> </math></EquationSource> </InlineEquation>) of gypseous sandy soil based on key soil parameters. An Extra Tree Regression (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$ETR$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">ETR</mi> </mrow> </math></EquationSource> </InlineEquation>) scheme was developed utilizing a database of 180 experimental records compiled from existing literature. To enhance model performance, the <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$ETR$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">ETR</mi> </mrow> </math></EquationSource> </InlineEquation>’s hyperparameters were optimized using two metaheuristic algorithms: Sea Horse Optimization (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">$SH$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">SH</mi> </mrow> </math></EquationSource> </InlineEquation>) and Artificial Protozoa Optimization (<InlineEquation ID="IEq6"> <EquationSource Format="TEX">$AP$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">AP</mi> </mrow> </math></EquationSource> </InlineEquation>). Seven input variables were used to estimate <InlineEquation ID="IEq7"> <EquationSource Format="TEX">$CP$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">CP</mi> </mrow> </math></EquationSource> </InlineEquation>. The results demonstrate that the recommended <InlineEquation ID="IEq8"> <EquationSource Format="TEX">$ML$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">ML</mi> </mrow> </math></EquationSource> </InlineEquation>-based approach offers a reliable framework for CP projection of gypseous sandy soils. The seven input elements are the following: first dry unit weights, first voids ratio, starting water content, specific gravity, gypsum content, and % passing sieve #200. Based on the results, both <InlineEquation ID="IEq9"> <EquationSource Format="TEX">$SH(ETR)$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>S</mi> <mi>H</mi> <mo stretchy="false">(</mo> <mi>E</mi> <mi>T</mi> <mi>R</mi> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq10"> <EquationSource Format="TEX">$AP(ETR)$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <mi>P</mi> <mo stretchy="false">(</mo> <mi>E</mi> <mi>T</mi> <mi>R</mi> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> demonstrated strong predictive capabilities for estimating <InlineEquation ID="IEq11"> <EquationSource Format="TEX">$CP$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">CP</mi> </mrow> </math></EquationSource> </InlineEquation>. The <InlineEquation ID="IEq12"> <EquationSource Format="TEX">$AP(ETR)$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <mi>P</mi> <mo stretchy="false">(</mo> <mi>E</mi> <mi>T</mi> <mi>R</mi> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> model acquired the peak R2, with values of 0.9901 for training and 0.9788 for testing. In contrast, the <InlineEquation ID="IEq13"> <EquationSource Format="TEX">$SH(ETR)$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>S</mi> <mi>H</mi> <mo stretchy="false">(</mo> <mi>E</mi> <mi>T</mi> <mi>R</mi> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> model produced R2 values of 0.9818 (training) and 0.9695 (testing), which were accompanied by higher error percentages-0.8396% during training and 0.9545% during testing-indicating lower reliability. Overall, the <InlineEquation ID="IEq14"> <EquationSource Format="TEX">$AP(ETR$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <mi>P</mi> <mo stretchy="false">(</mo> <mi>E</mi> <mi>T</mi> <mi>R</mi> </mrow> </math></EquationSource> </InlineEquation>) approach outperformed <InlineEquation ID="IEq15"> <EquationSource Format="TEX">$SH(ETR)$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>S</mi> <mi>H</mi> <mo stretchy="false">(</mo> <mi>E</mi> <mi>T</mi> <mi>R</mi> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation>, offering more accurate and robust predictions.

Collapse capacity of the gypseous sandy soils: usefulness of extra tree regressions

期刊论文
发表日期：

2025年12月19日

查看全文

Daoyong Zhu,
Qin Yan,
Xiangzhuo Yu,
Xiao Wu

摘要

This exploration reviews machine learning ( $ML$ $ML$ ) tactics to project the collapse potential ( $CP$ $CP$ ) of gypseous sandy soil based on key soil parameters. An Extra Tree Regression ( $ETR$ $ETR$ ) scheme was developed utilizing a database of 180 experimental records compiled from existing literature. To enhance model performance, the $ETR$ $ETR$ ’s hyperparameters were optimized using two metaheuristic algorithms: Sea Horse Optimization ( $SH$ $SH$ ) and Artificial Protozoa Optimization ( $AP$ $AP$ ). Seven input variables were used to estimate $CP$ $CP$ . The results demonstrate that the recommended $ML$ $ML$ -based approach offers a reliable framework for CP projection of gypseous sandy soils. The seven input elements are the following: first dry unit weights, first voids ratio, starting water content, specific gravity, gypsum content, and % passing sieve #200. Based on the results, both $SH(ETR)$ $S H (E T R)$ and $AP(ETR)$ $A P (E T R)$ demonstrated strong predictive capabilities for estimating $CP$ $CP$ . The $AP(ETR)$ $A P (E T R)$ model acquired the peak R², with values of 0.9901 for training and 0.9788 for testing. In contrast, the $SH(ETR)$ $S H (E T R)$ model produced R² values of 0.9818 (training) and 0.9695 (testing), which were accompanied by higher error percentages-0.8396% during training and 0.9545% during testing-indicating lower reliability. Overall, the $AP(ETR$ $A P (E T R$ ) approach outperformed $SH(ETR)$ $S H (E T R)$ , offering more accurate and robust predictions.