Scholarlink

<p>In order to create environmentally friendly building materials, a novel approach involves incorporating carbon-rich particles that are produced by burning biomass into cementitious mixtures. By doing so, resource circularity can be improved, clinker consumption can be reduced, and stable biogenic carbon can be stored in infrastructure that is designed to persist for a long time. But the effectiveness of this material is contingent on a number of factors, including the amount of biochar that is combined with concrete, the amount of biomass that is utilized, the temperature at which it is broken down, the size of the particles, the minerals that are utilized, the length of time that they cure, and the structure of the pores. Due to the presence of these factors, behavior does not follow a straight line, which is something that cannot be clearly demonstrated by typical observational strength correlations on their own. In an effort to find a solution to this issue, we have developed a multi-model machine learning system that is capable of estimating the compressive strength, splitting tensile strength, flexural strength, and elastic modulus of the biochar-concrete combined material. The following are some of the components that are utilized in this system: filtered experimental and research data, feedstock-mineral descriptors, autoencoder-based feature compression, Bayesian uncertainty estimation, stacked ensemble learning, particle swarm optimization, and SHAP-based interpretation process. The model is not intended to fulfill the role of a black-box regression tool; rather, it is intended to assist with correct predictions as well as the comprehension of contents. In summary, such effects are used within dimensionality reduction with autoencoders and feature engineering with extraction of latent variables from high-dimensional data of pyrolysis, ensuring reduction in dimensions to 50–70% along with greater reconstruction accuracy to 90%. Next, using Bayesian neural networks (BNNs, ) framework managed the quantification of uncertainty such that the given below predictions on the mechanical properties lay within the range of a 95% confidence interval with mean prediction error to below 10%. Transfer learning with pre-trained convolutional neural networks (CNNs) decreased the training time by 30% and increased accuracy, with the prediction of compressive strength within 5% of experimental values. Another ensemble model, the method of stacked generalization combined stochastic forests with deep neural networks and resulted in an additional 15–20% error reduction in the prediction. The application of particle swarm optimization (PSO) in the hyperparameter tuning phase improved the accuracy by 10%, while the training time decreased by 25%. Finally, Shapley additive explanations (SHAP) attributed model explainability and showed that pyrolysis temperature and particle size were very influencing factors on the predictions. The primary findings indicate that the addition of biochar in quantities ranging from low to moderate is beneficial to both the predictive and mechanical reactions, provided that the dosage, feedstock, and pyrolysis parameters are maintained in the appropriate manner. In the independent test set, it had a mean absolute error of 2.31&#xa0;MPa and a coefficient of determination of 0.94 for compressive strength, 0.42&#xa0;MPa and 0.91 for flexural strength, 0.29&#xa0;MPa and 0.90 for splitting tensile strength, and 1.35 GPa and 0.89 for elastic modulus. It is a very useful tool for optimizing biochar infused concrete on sustainable construction application grounds because of its completeness; it enhances the accuracy in prediction, reduces uncertainty.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimization and prediction of mechanical properties of biochar-sustainable concrete modified with agro-industrial by-products residue feedstock

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月22日

查看全文
  • Vikrant S. Vairagade

摘要

In order to create environmentally friendly building materials, a novel approach involves incorporating carbon-rich particles that are produced by burning biomass into cementitious mixtures. By doing so, resource circularity can be improved, clinker consumption can be reduced, and stable biogenic carbon can be stored in infrastructure that is designed to persist for a long time. But the effectiveness of this material is contingent on a number of factors, including the amount of biochar that is combined with concrete, the amount of biomass that is utilized, the temperature at which it is broken down, the size of the particles, the minerals that are utilized, the length of time that they cure, and the structure of the pores. Due to the presence of these factors, behavior does not follow a straight line, which is something that cannot be clearly demonstrated by typical observational strength correlations on their own. In an effort to find a solution to this issue, we have developed a multi-model machine learning system that is capable of estimating the compressive strength, splitting tensile strength, flexural strength, and elastic modulus of the biochar-concrete combined material. The following are some of the components that are utilized in this system: filtered experimental and research data, feedstock-mineral descriptors, autoencoder-based feature compression, Bayesian uncertainty estimation, stacked ensemble learning, particle swarm optimization, and SHAP-based interpretation process. The model is not intended to fulfill the role of a black-box regression tool; rather, it is intended to assist with correct predictions as well as the comprehension of contents. In summary, such effects are used within dimensionality reduction with autoencoders and feature engineering with extraction of latent variables from high-dimensional data of pyrolysis, ensuring reduction in dimensions to 50–70% along with greater reconstruction accuracy to 90%. Next, using Bayesian neural networks (BNNs, ) framework managed the quantification of uncertainty such that the given below predictions on the mechanical properties lay within the range of a 95% confidence interval with mean prediction error to below 10%. Transfer learning with pre-trained convolutional neural networks (CNNs) decreased the training time by 30% and increased accuracy, with the prediction of compressive strength within 5% of experimental values. Another ensemble model, the method of stacked generalization combined stochastic forests with deep neural networks and resulted in an additional 15–20% error reduction in the prediction. The application of particle swarm optimization (PSO) in the hyperparameter tuning phase improved the accuracy by 10%, while the training time decreased by 25%. Finally, Shapley additive explanations (SHAP) attributed model explainability and showed that pyrolysis temperature and particle size were very influencing factors on the predictions. The primary findings indicate that the addition of biochar in quantities ranging from low to moderate is beneficial to both the predictive and mechanical reactions, provided that the dosage, feedstock, and pyrolysis parameters are maintained in the appropriate manner. In the independent test set, it had a mean absolute error of 2.31 MPa and a coefficient of determination of 0.94 for compressive strength, 0.42 MPa and 0.91 for flexural strength, 0.29 MPa and 0.90 for splitting tensile strength, and 1.35 GPa and 0.89 for elastic modulus. It is a very useful tool for optimizing biochar infused concrete on sustainable construction application grounds because of its completeness; it enhances the accuracy in prediction, reduces uncertainty.