Machine learning maps biochar’s strength effects in cement, with limits

Biochar has become one of the most discussed candidates for greening cement-based construction materials, but a new study suggests the scientific literature on the subject is still too fragmented to support reliable predictions of how the carbon-rich additive will perform in any given mix. Researchers have compiled a structured database from published studies on biochar-modified cementitious composites and subjected it to machine learning analysis, with results that are as revealing about the state of the literature as they are about the material itself.

The reasoning behind this normalisation is technical but important. Absolute compressive strength in cementitious materials is dominated by the baseline matrix. A biochar-modified ultra-high-performance concrete and a biochar-modified brick occupy entirely different strength ranges before any biochar is added, so modelling absolute strength directly risks teaching a model to distinguish material classes rather than to capture the relative influence of the additive.

From the 56 sources, the team extracted 368 rows of data. After cleaning and excluding 22 control rows that equal unity by definition, the final modelling dataset contained 131 biochar-modified rows. Only 15 of the 56 sources ultimately contributed rows sufficiently complete for the machine learning subset, a finding the authors describe as one of the more informative results of the database audit.

Under a conventional random 75/25 train-test split, the Tier 1 ridge regression model achieved a coefficient of determination of 0.335, with a mean absolute error of 0.118 and a root mean squared error of 0.164 in strength-ratio units. Those numbers suggest the selected predictors contained useful internal information within the compiled database.

Under source-based validation, performance collapsed. The best Tier 1 result came from random forest, with a coefficient of determination of minus 0.030, a mean absolute error of 0.170 and a root mean squared error of 0.207. A negative coefficient of determination means the model performed worse than a simple mean-based prediction for unseen sources. Nested group-aware hyperparameter tuning did not restore positive Tier 1 performance, with the best tuned model reaching only minus 0.008. The authors attribute the gap between the two validation strategies to source-specific patterns: rows from the same publication often share raw materials, mix-design logic, curing conditions and reporting style, so random splitting can let a model benefit from patterns it has effectively already seen.

The tuned Tier 1 random forest applied to this external data produced a coefficient of determination of minus 0.153, a mean absolute error of 0.082 and a root mean squared error of 0.108. Predictions were positively biased by 0.070 strength-ratio units on average. Yet the model showed a strong rank correlation with observations, with a Spearman coefficient of 0.797, indicating partial recognition of trends without reliable calibration.

On the question of which variables matter most, biochar dosage emerged as the leading predictor in both impurity-based feature importance and grouped SHAP analysis, contributing roughly 47.6 percent of aggregated importance in the former and 30.5 percent in the latter. Other influential predictors included pyrolysis temperature, water absorption, control compressive strength, particle size, biochar type, material type and curing method. The authors caution that these rankings are database-derived associations rather than causal evidence, particularly because several mechanistically important variables were sparsely reported and required imputation. Water absorption, for instance, appeared in fewer than one-quarter of rows.

The statistical audit of the database also quantified how much of the literature’s variability stems from the baseline matrix rather than biochar itself. Material type showed a large association with absolute biochar-mixture strength, but after conversion to strength ratio the association was reduced, though not eliminated. Absolute strength correlated very strongly with control strength, with a Spearman coefficient of 0.937, while the strength-ratio correlation was much weaker at 0.291. Normalisation substantially reduced the dominance of baseline matrix strength but left residual heterogeneity. Within the compiled data, strength ratios ranged from 0.398 to 1.616, with a mean of 0.963 and a median of 0.985, and roughly 22 percent of rows fell between 0.95 and 1.05, indicating performance close to the corresponding controls.

The authors are explicit that the framework should be understood as exploratory, diagnostic and source-aware rather than as a universal or deployable prediction model. They identify several structural limitations: the dataset is modest in size and source-imbalanced, with the five most represented studies contributing approximately 61 percent of usable rows; it combines cementitious systems that differ substantially in binder composition, aggregate structure and baseline strength; and key predictors such as water-to-binder ratio, particle size, BET surface area, pH, carbon content and ash content were frequently missing. Imputation allowed the pipeline to run but should not be interpreted as measured material properties. The practical value of the work, they argue, lies in organising a fragmented literature base, assessing model transferability, identifying influential variables and making the reporting gaps that limit future predictive modelling more visible.

Subject of Research: Machine learning analysis of a literature-derived database on compressive strength retention in biochar-modified cementitious composites

Article Title: A literature-derived database and machine learning-assisted analysis of compressive strength retention in biochar-modified cementitious composites

Article References: Osman, S., Mokhtar, S. H., Zakaria, M. N., Bakar, B. F. A., & Wan Jaafar, W. N. R. (2026). A literature-derived database and machine learning-assisted analysis of compressive strength retention in biochar-modified cementitious composites. Discover Concrete and Cement, 2(1), Article 75. https://doi.org/10.1007/s44416-026-00118-9

Image Credits: AI Generated

DOI: 10.1007/s44416-026-00118-9

Keywords: biochar, cementitious composites, compressive strength, machine learning, strength ratio, leave-one-source-out validation, random forest, SHAP, construction materials, data reporting gaps, cement replacement, sustainable building