<p>Low hydration activity and challenging management of delayed expansion restrict the application of fly ash <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\left(FA\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <mi>F</mi> <mi>A</mi> </mfenced> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$MgO$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">MgO</mi> </mrow> </math></EquationSource> </InlineEquation> expansive additive <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\left(MEA\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <mi>M</mi> <mi>E</mi> <mi>A</mi> </mfenced> </math></EquationSource> </InlineEquation> in cement pastes. Very few research on machine learning techniques for volume expansion <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$\left({V}_{e}\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <msub> <mi>V</mi> <mi>e</mi> </msub> </mfenced> </math></EquationSource> </InlineEquation> in such systems have been conducted already. This study develops a machine learning framework to predict the volume expansion <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$\left({V}_{e}\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <msub> <mi>V</mi> <mi>e</mi> </msub> </mfenced> </math></EquationSource> </InlineEquation> of cement pastes incorporating fly ash <InlineEquation ID="IEq6"> <EquationSource Format="TEX">$\left(FA\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <mi>F</mi> <mi>A</mi> </mfenced> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq7"> <EquationSource Format="TEX">$MgO$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">MgO</mi> </mrow> </math></EquationSource> </InlineEquation> expansive additive <InlineEquation ID="IEq8"> <EquationSource Format="TEX">$\left(MEA\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <mi>M</mi> <mi>E</mi> <mi>A</mi> </mfenced> </math></EquationSource> </InlineEquation>, materials whose low hydration activity and delayed expansion complicate their practical use. A dataset of 170 samples compiled from published literature was utilized, comprising four input variables—Portland cement content (<InlineEquation ID="IEq9"> <EquationSource Format="TEX">$PC$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">PC</mi> </mrow> </math></EquationSource> </InlineEquation>, %), fly ash content (<InlineEquation ID="IEq10"> <EquationSource Format="TEX">$FA$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">FA</mi> </mrow> </math></EquationSource> </InlineEquation>, %), <InlineEquation ID="IEq11"> <EquationSource Format="TEX">$MgO$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">MgO</mi> </mrow> </math></EquationSource> </InlineEquation> expansive additive (<InlineEquation ID="IEq12"> <EquationSource Format="TEX">$MEA$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">MEA</mi> </mrow> </math></EquationSource> </InlineEquation>, %), and curing age (<InlineEquation ID="IEq13"> <EquationSource Format="TEX">$SA$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">SA</mi> </mrow> </math></EquationSource> </InlineEquation>, days)—to predict the target variable, <InlineEquation ID="IEq14"> <EquationSource Format="TEX">${V}_{e}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>V</mi> <mi>e</mi> </msub> </math></EquationSource> </InlineEquation>(%). A Categorical Boosting <InlineEquation ID="IEq15"> <EquationSource Format="TEX">$\left(CatBoost\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <mi>C</mi> <mi>a</mi> <mi>t</mi> <mi>B</mi> <mi>o</mi> <mi>o</mi> <mi>s</mi> <mi>t</mi> </mfenced> </math></EquationSource> </InlineEquation> model was optimized using two recent metaheuristic algorithms, the Starfish Optimization Algorithm <InlineEquation ID="IEq16"> <EquationSource Format="TEX">$\left(StOA\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <mi>S</mi> <mi>t</mi> <mi>O</mi> <mi>A</mi> </mfenced> </math></EquationSource> </InlineEquation> and the Flood Optimization Algorithm <InlineEquation ID="IEq17"> <EquationSource Format="TEX">$\left(FlOA\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <mi>F</mi> <mi>l</mi> <mi>O</mi> <mi>A</mi> </mfenced> </math></EquationSource> </InlineEquation>, to enhance prediction accuracy. Model performance was assessed through cross-validation, normalization, and feature importance analysis, with 70% of the data (n = 119) used for training and 30% (n = 51) for testing. The optimized <InlineEquation ID="IEq18"> <EquationSource Format="TEX">$CatBoost-StOA$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>C</mi> <mi>a</mi> <mi>t</mi> <mi>B</mi> <mi>o</mi> <mi>o</mi> <mi>s</mi> <mi>t</mi> <mo>-</mo> <mi>S</mi> <mi>t</mi> <mi>O</mi> <mi>A</mi> </mrow> </math></EquationSource> </InlineEquation> model achieved the best performance, with <InlineEquation ID="IEq19"> <EquationSource Format="TEX">$RMSLE$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="italic">RMSLE</mi> </mrow> </math></EquationSource> </InlineEquation> values of 0.00404 (training) and 0.0053 (testing), corresponding to low mean absolute and root mean square errors in predicting <InlineEquation ID="IEq20"> <EquationSource Format="TEX">${V}_{e}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>V</mi> <mi>e</mi> </msub> </math></EquationSource> </InlineEquation>. These results demonstrate the model’s potential as a reliable predictive tool for mix design optimization and dimensional stability control in cementitious materials.</p>

Enhancing dimensional stability in cement pastes: a machine learning approach with metaheuristic-optimized categorical boosting

期刊论文
出版模式：

开放获取
发表日期：

2026年2月12日

查看全文

Mahzad Esmaeili-Falak,
Behzad Sojoudi

摘要

Low hydration activity and challenging management of delayed expansion restrict the application of fly ash $\left(FA\right)$ $(F, A)$ and $MgO$ $MgO$ expansive additive $\left(MEA\right)$ $(M, E, A)$ in cement pastes. Very few research on machine learning techniques for volume expansion $\left({V}_{e}\right)$ $(V_{e})$ in such systems have been conducted already. This study develops a machine learning framework to predict the volume expansion $\left({V}_{e}\right)$ $(V_{e})$ of cement pastes incorporating fly ash $\left(FA\right)$ $(F, A)$ and $MgO$ $MgO$ expansive additive $\left(MEA\right)$ $(M, E, A)$ , materials whose low hydration activity and delayed expansion complicate their practical use. A dataset of 170 samples compiled from published literature was utilized, comprising four input variables—Portland cement content ( $PC$ $PC$ , %), fly ash content ( $FA$ $FA$ , %), $MgO$ $MgO$ expansive additive ( $MEA$ $MEA$ , %), and curing age ( $SA$ $SA$ , days)—to predict the target variable, ${V}_{e}$ $V_{e}$ (%). A Categorical Boosting $\left(CatBoost\right)$ $(C, a, t, B, o, o, s, t)$ model was optimized using two recent metaheuristic algorithms, the Starfish Optimization Algorithm $\left(StOA\right)$ $(S, t, O, A)$ and the Flood Optimization Algorithm $\left(FlOA\right)$ $(F, l, O, A)$ , to enhance prediction accuracy. Model performance was assessed through cross-validation, normalization, and feature importance analysis, with 70% of the data (n = 119) used for training and 30% (n = 51) for testing. The optimized $CatBoost-StOA$ $C a t B o o s t - S t O A$ model achieved the best performance, with $RMSLE$ $RMSLE$ values of 0.00404 (training) and 0.0053 (testing), corresponding to low mean absolute and root mean square errors in predicting ${V}_{e}$ $V_{e}$ . These results demonstrate the model’s potential as a reliable predictive tool for mix design optimization and dimensional stability control in cementitious materials.