Work

Artificial-intelligence-assisted inverse design of low-carbon concrete under carbonation constraints

Year: 2026

Type: article

Abstract: This study develops and tests an artificial-intelligence-assisted inverse-design workflow for supplementary-cementitious-material-blended concrete, with artificial intelligence applied to surrogate prediction, uncertainty quantification, dosage-effect interpretation, and multi-objective candidate generation. Optuna-tuned feedforward neural networks were trained on four open datasets and coupled with split conformal prediction, continuous-treatment forest analysis, and non-dominated multi-objective search. The source-aware compressive-strength model achieved a mean absolute error of 7.20 megapascals and a coefficient of determination of 0.687. Under standard cross-validation, embodied carbon achieved a mean absolute error of 2.41 and a coefficient of determination of 0.999, while carbonation depth achieved a mean absolute error of 2.56 millimetres and a coefficient of determination of 0.847. A leakage audit showed that the exceptional embodied-carbon fit is largely attributable to an almost deterministic relation between mixture quantities and the carbon target, although performance was sensitive to source grouping. Continuous-treatment re-analysis showed substantial pre-adjustment imbalance and wide ninety-five-percent effect intervals, so the dosage results are interpreted conditionally rather than as universal causal effects. Three-objective optimization produced seventy-two computational candidates; sixteen, or 22.2 percent, had negative raw carbonation predictions and were excluded from representative reporting. A physically non-negative balanced candidate predicted 75.8 megapascals strength, 141.7 embodied carbon, and 0.064 millimetres carbonation depth. Applicability-domain analysis showed that none of the optimized candidates lay inside a strict joint ninety-five-percent domain for all retained endpoints, emphasizing that the generated mixtures are hypotheses for laboratory validation rather than design-ready prescriptions.

Source: Engineering Applications of Artificial Intelligence

DOI: 10.1016/j.engappai.2026.116633

Authors: Lenganji Simwanda, Paul O. Awoyera, Adewumi John Babafemi

Institutions: Northrise University, Czech Technical University in Prague, Prince Mohammad bin Fahd University, Stellenbosch University

Language: English

Topic: Concrete and Cement Materials Research

Subfield: Civil and Structural Engineering

Field: Engineering

Domain: Physical Sciences

SDGs: Industry, innovation and infrastructure, Climate action

Open Access status: hybrid

Estimated APC paid: $3,200