<scp>An</scp> interpretable tabular meta‐learning method for predicting the compressive strength of glass powder concrete

Abstract Glass powder (GP) can partially replace cement and promote the resource utilization of waste glass, but its effect on compressive strength is governed by multiple coupled factors, making cross‐literature prediction uncertain. This study developed an interpretable TabPFN‐based tabular meta‐learning framework for predicting the compressive strength of ordinary GP concrete. A database of 728 records from 29 peer‐reviewed studies was compiled, with all mixtures restricted to GP as the sole partial cement replacement. Input variables included mix proportions, curing age, admixtures, and major oxide composition. The effects of GP replacement level, model accuracy, robustness, and variable influence patterns were evaluated using the relative strength ratio (RSR), 10‐fold cross‐validation, 2000 Monte Carlo random splits, and SHAP‐GAM analysis. The median RSR values at 5% and 10% replacement levels were close to or slightly above 1, with RSR>1 proportions of approximately 67% and 52%, respectively; this proportion decreased below 50% when the replacement level reached 15% or higher. TabPFN achieved the best performance, with MAE, RMSE, MAPE, and R 2 values of 1.5787 MPa, 2.3465 MPa, 6.6399%, and 0.9625, respectively. Compared with CatBoost, MAE and RMSE were reduced by approximately 12.0% and 10.9%. SHAP‐GAM analysis identified curing age and water‐to‐binder ratio (W/B) as the dominant variables, with critical thresholds at a W/B of approximately 0.49 and a GP dosage of approximately 105.77 kg/m 3 . The proposed framework provides interpretable data‐driven support for mixture optimization, quality control, and engineering applicability assessment of GP concrete.