The research, led by M. Vignesh, S.S. Alan Prince, Christina Elangbam and Vasugi K, arrives at a moment when the concrete industry is undergoing a profound compositional transformation. To cut the enormous carbon footprint of cement production, manufacturers increasingly replace a portion of cement with supplementary cementitious materials, or SCMs, such as fly ash, silica fume and ground-granulated blast-furnace slag. These industrial by-products lower CO2 emissions while improving long-term strength and durability. Chemical admixtures, including high-range water reducers and accelerating agents, have expanded the design space even further, allowing engineers to push water-to-binder ratios lower and achieve strength faster without sacrificing workability. The trouble is that the classical empirical tools of the trade, from the Abrams water-cement law to the ACI 211.1 mix-design charts and the prescriptive provisions of EN 206, were calibrated for simpler binder systems. With ternary and even quaternary blends now common, mixes routinely fall outside the calibration range of the old equations.

Machine learning has long promised a way out, and neural networks, support vector regression and tree-based ensembles have all been applied to concrete strength prediction, with reported coefficients of determination on held-out test sets approaching or exceeding 0.90. But three persistent problems have kept these methods out of everyday engineering practice. Most published studies rely on small, single-laboratory datasets of uncertain provenance; validation typically rests on a single train-test split rather than rigorous cross-validation and residual diagnostics; and the most accurate models, gradient-boosted ensembles, are widely regarded as black boxes that offer no transparent account of why they predict what they predict. Before structural engineers and code committees can trust an algorithm’s output, they need physically justifiable relationships between inputs and outputs. The new study tackles all three shortcomings at once.

The foundation of the work is a dataset of 1,670 mix records drawn from the field ledger of the Guangxi Road and Bridge Group’s Road and Bridge Pavement Branch, deposited via Mendeley Data. Because the records come from normal construction practice rather than a controlled laboratory programme, they capture strength outcomes as they actually occur on site, including the effects of curing conditions, compaction and specimen handling. Each record contains nine measured variables: cement, fine aggregates, coarse aggregates, water, water-reducing admixture, fly ash, accelerating agent, silica fume and curing time. The target compressive strength spans a remarkable 4.3 to 76.3 megapascals, with a mean of 37.62 megapascals, covering everything from low-grade formulations to high-performance mixes. Curing ages cluster bimodally at 7 and 28 days, which allowed the model to learn age-dependent strength gain in a single training exercise. Crucially, no records were discarded and no outlier removal was performed, preserving the messiness of reality.

On top of the nine raw variables, the team engineered three physics-informed features drawn directly from classical concrete theory. The water-cement ratio encodes the Abrams law, sparing the algorithm from having to rediscover the single most famous relationship in concrete technology. The cement ratio, the binder’s share of the total aggregate skeleton, governs paste pressure and the density of the interfacial transition zone where strength-critical failure often begins. And the total cementitious content, the sum of cement, fly ash and silica fume, captures the full pool of hydration-active material available for long-term strength development. The final feature matrix thus held twelve predictors, and all were standardised using parameters fitted exclusively to the training data to prevent information leakage, with the strength target left unscaled so that error metrics remained directly interpretable in megapascals.

Six regression algorithms were then benchmarked on identical data partitions with a fixed random seed: linear regression, k-nearest neighbours, support vector regression with a radial-basis-function kernel, random forest, LightGBM and XGBoost. The hierarchy that emerged was textbook-clear. Linear regression managed only a coefficient of determination of 0.7755 with a root-mean-square error of 5.81 megapascals, its ordinary least-squares machinery simply unable to capture the nonlinear, interaction-dominated response surface. Support vector regression and k-nearest neighbours landed in the mid-0.83 range, while random forest crossed the 0.90 engineering-grade threshold. LightGBM pushed to 0.9209, and XGBoost led the untuned field at 0.9243 with a root-mean-square error of 3.37 megapascals. Ten-fold cross-validation confirmed the ranking, with XGBoost achieving a mean coefficient of determination of 0.9127 plus or minus 0.0318 across folds, evidence that the accuracy reflected genuine learned trends rather than a lucky test-set composition.

Hyperparameter optimisation via a five-fold grid search over 27 combinations, 135 model fits in total, nudged XGBoost to a final held-out performance of R-squared 0.9263, root-mean-square error 3.329 megapascals and mean absolute error 2.40 megapascals. The optimal configuration, 300 estimators, a maximum tree depth of 4 and a learning rate of 0.1, favoured shallower trees with faster learning than the baseline, consistent with the known behaviour of gradient boosting on moderately sized datasets. Residual diagnostics showed errors centred on zero across the entire 4-to-70-megapascal prediction range, with an essentially Gaussian residual distribution and no directional bias. Of the 334 test samples, 76 percent fell within 10 percent of the measured strength and 92.8 percent within 20 percent. Range-wise analysis added a note of caution: in the narrow 40-to-55-megapascal band the coefficient of determination dropped sharply, though absolute errors stayed low, a reminder that R-squared must always be read alongside error metrics and sample counts.

The real novelty lies in what came next: a three-layer explainability analysis. Global feature importance computed across four model families converged on the same ranking, with the water-cement ratio leading in both random forest and XGBoost, followed by total cementitious content, curing time and water. SHAP values, rooted in cooperative game theory, then quantified each feature’s contribution for every individual prediction. The water-cement ratio produced SHAP values spanning from minus 25 to plus 15 megapascals, the widest influence of any predictor: low ratios pushed predictions well above the 37.87-megapascal baseline while high ratios imposed penalties of 10 to 15 megapascals, a modern, multi-constituent confirmation of the century-old Abrams law. Curing time contributed negatively for 7-day specimens and positively for 28-day ones, and partial-dependence plots traced a near-linear strength gain from roughly 33.5 megapascals at 7 days to 40.5 megapascals at 28 days. The water partial-dependence curve revealed a strikingly nonlinear, monotonically declining response consistent with the classical gel-space ratio concept of paste porosity.

Individual waterfall decompositions show the practical payoff. For one weak mix that actually achieved 22.6 megapascals, the model predicted 26.7, and the explanation was legible in engineering terms: total cementitious content subtracted 4.16 megapascals from the baseline, the water-cement ratio another 4.00, short curing 3.28, and an unfavourable fly-ash dosage 1.27 more. The authors are careful to interpret such attributions correctly, noting that the sparse, zero-inflated field data on fly ash and silica fume, present in only 34.7 percent and 1.1 percent of records respectively, mean the model captures directional pozzolanic sensitivity rather than a validated dose-response curve. Independent microstructural studies, which found strength optima near 30 percent fly ash and 7.5 to 12 percent silica fume replacement, provide complementary mechanistic context. The team is equally candid about limitations: no new laboratory programme was conducted, so validation is statistical rather than experimental, and the web-based implementation is a proof-of-concept decision-support tool, not a replacement for standard strength testing. Still, by fusing transparent field data, physics-informed feature engineering, systematic benchmarking and multi-level explainability into one reproducible framework, the study sketches what trustworthy artificial intelligence for concrete design could look like, an algorithm that not only predicts strength within about three megapascals but can show its work, one mix at a time.

Subject of Research: Explainable machine learning prediction of compressive strength in SCM-enhanced concrete at the mix-design stage

Article Title: Explainable machine learning for mix-design-stage prediction of compressive strength in SCM-enhanced concrete

Article References: Explainable machine learning for mix-design-stage prediction of compressive strength in SCM-enhanced concrete. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: concrete, compressive strength, machine learning, XGBoost, SHAP, supplementary cementitious materials, fly ash, silica fume, water-cement ratio, explainable AI, mix design, construction materials