Cement production contributes a significant share of global CO \(_2\) emissions, yet most laboratories can only access small concrete mix datasets, which limits the direct use of data-driven design tools. This study develops a physics-guided deep ensemble and inverse design framework to support the design of lower-carbon concrete mixtures under data scarcity. The framework first augments a 103-record concrete slump dataset with a physics-regularized conditional tabular GAN (PR-CTGAN) that enforces mass balance, water–binder bounds, and admixture dosage limits during synthetic data generation. It then trains a heterogeneous deep ensemble that combines tree-based regressors with a deep evidential regression (DER) network and a physics-regularized neural network (PRNN) that encodes an empirical slump–water–binder relation as a soft penalty in the loss. This ensemble predicts slump, flow, and 28-day compressive strength while providing uncertainty estimates for each target. Multi-objective Bayesian optimisation tunes the evidential backbone to balance accuracy and probabilistic calibration, and explainable AI tools (SHAP and Sobol sensitivity analysis) highlight how water, binder chemistry, and aggregate ratios drive fresh and hardened behaviour in a way that aligns with concrete practice. Finally, an NSGA-III-based inverse design stage searches the mix space for candidate formulations that meet workability and strength targets while lowering estimated binder-related CO \(_2\) emissions compared with an all-cement reference mix. The framework integrates physics-guided data augmentation, uncertainty-aware evidential prediction, and eco-constrained inverse optimisation into a single pipeline for sustainable concrete mix design.
