Physics-Guided Generative Inverse Design of Low-Carbon Concrete via Multi-Objective Latent Space Optimization

Designing low-carbon concrete with high mechanical performance, low environmental impact, and acceptable cost is a challenging multiobjective optimization problem. This study develops an integrated generative inverse-design framework that combines Variational Autoencoder (VAE)-based latent-space representation, a physics-guided strength surrogate model, and non-dominated sorting genetic algorithm II (NSGA-II) optimization. A database containing 2371 concrete mixture records was constructed from public data sets and processed through unit harmonization, duplicate checking, missing-value treatment, and physical-feasibility screening. The VAE was used to learn a continuous latent representation of feasible concrete mixture patterns, while the physics-guided strength surrogate predicted compressive strength using regularization terms related to water-to-binder ratio, age-dependent strength development, and mixture feasibility. Material-related cradle-to-gate GWP and material cost were directly calculated from decoded mixture proportions using cradle-to-gate emission factors and unit prices. NSGA-II was then conducted in the latent space to generate Pareto-optimal solutions balancing strength, GWP, and cost. Three representative mixtures selected from distinct regions of the Pareto Frontierecological optimum, cost optimum, and balanced solutionwere experimentally prepared and tested as a preliminary laboratory check. The measured 28 day compressive strengths showed good agreement with model predictions, with relative errors below 5% for these selected Pareto solutions. These results provide preliminary evidence that the proposed workflow can generate feasible candidate mixtures under the adopted strength, GWP, cost, and constraint settings. Further validation using additional mixtures and broader performance indicators, including workability, early age strength, and durability, is required before general practical applicability can be established.