A model agnostic procedural framework for reverse concrete mix design using existing artificial intelligence property predictors

Artificial intelligence models in concrete materials research usually operate as forward property predictors. They receive mixture variables and curing information, then return an estimated material property. Concrete proportioning, however, normally begins with a required performance condition and proceeds toward a feasible mixture. The gap addressed in this study is the absence of a practical, auditable, model-agnostic protocol for using an existing or fixed forward property predictor within a target-first reverse-design workflow, particularly when the predictor’s development information is incomplete or is not controlled by the end user. This study develops a procedural framework for reverse concrete mix design with existing artificial intelligence property predictors and treats the predictor as a virtual trial evaluator inside a human-guided design procedure. The framework defines stages for target translation, model eligibility screening, candidate feasibility, reverse trial adjustment, uncertainty-aware acceptance, engineering coherence checking, and reproducible reporting. A numerical case study using a 3544-record concrete mixture dataset trained an extremely randomized trees predictor and applied the proposed gates to a 40 MPa, 28-day reverse design target. The predictor achieved test R 2 = 0.972, RMSE = 6.63 MPa, and MAE = 4.18 MPa; the calibration-based uncertainty allowance was 8.25 MPa. Using the conservative score, 42 of 149 held-out feasible candidates were accepted, and 40 of those accepted candidates had measured strength not less than 40 MPa. The selected minimum-binder candidate had predicted strength 51.68 MPa and measured strength 49.30 MPa. The contribution is procedural and validation-focused rather than a new optimization algorithm, and the final AI-assisted candidate remains subject to laboratory and specification verification.