Conventional development of low-carbon clinker mainly relies on empirical or mechanistic understanding of high-temperature and hydration reactions among clinker phases. This strategy, which combines experimental trial-and-error with mechanistic analysis, is generally time-consuming. Therefore, conventional development approaches are no longer sufficient to meet the demand for the rapid development of novel low-carbon clinkers. The development of high-throughput (HT) methods and machine learning (ML) has provided a new paradigm for the design of cement-based materials. However, available clinker data for ML-based clinker design are usually collected from scattered sources, and they differ substantially in preparation procedures, testing conditions, and compositional definitions, making them difficult to use directly. To address these issues, this study used clinker phases as compositional units and constructed a clinker dataset through a standardized HT preparation procedure. On this basis, the predictive performance of different ML models for the 28 d compressive strength of clinker was compared. Furthermore, the optimized surrogate model was coupled with a genetic algorithm to screen for low-carbon clinkers under constraints on 28 d compressive strength and CO₂ emissions. This approach enabled evaluation of the applicability of the proposed data-driven framework for the design of novel low-carbon clinkers.

In this study, 11 clinker phases were selected as the basis for compositional design, and each clinker composition prepared by dry-mixing these phases with gypsum in specified proportions was regarded as one specimen. A clinker dataset was constructed using an HT experimental procedure. The dataset contained information on phase composition, water-to-binder ratio (W/B), 3 and 28 d compressive strengths for a total of 264 specimens. Among them, 228 specimens were used for training and validation, while 36 specimens were reserved as a test set. The test set was further divided into in-system and out-of-system specimens to evaluate model stability. For model construction, five nonlinear ML algorithms were employed. The normalized phase composition after gypsum incorporation was used as the input. The 28 d compressive strength was used as the output. Model hyperparameters were determined by Bayesian optimization, and repeated constrained holdout validation was adopted to reduce the influence of random splitting on model evaluation. Model performance was evaluated using mean squared error, root mean squared error (RMSE), mean absolute error, and mean absolute percentage error. After the surrogate model had been determined, it was coupled with a genetic algorithm to search the phase composition space for low-carbon clinkers with a predicted 28 d compressive strength higher than 42.5 MPa and CO₂ emissions lower than 671 or 592 kg CO₂/t. These CO₂-emission thresholds were reduced by 15% and 25%, respectively, relative to Portland clinker.

Dataset analysis showed that the specimens in the dataset exhibited pronounced compositional dispersion and system-level heterogeneity. This heterogeneity was one of the main sources of model instability. The W/B gradually increased from the C3S/C2S-rich region toward regions enriched in aluminate/ferrite phases and C4A3$/C5S2$. The 3 d and 28 d compressive strengths of the training/validation set covered a range of 0-70 MPa, whereas the test set was mainly concentrated in the medium-to-high strength range. Some specimens exhibited strength regression from 3 d to 28 d. These complex strength-development patterns increased the difficulty of model learning but also provided useful information for identifying nonlinear relationships between composition and compressive strength. Validation results showed that, under the conditions of limited data and high compositional heterogeneity, controlled model capacity and moderate regularization were more beneficial for improving prediction stability. The optimized hyperparameters of different models all reflected a tendency to suppress overfitting. Among the models, the artificial neural network (ANN) adopted a relatively small architecture with regularization and achieved the best performance on the validation sets. The test-set results further demonstrated that the ANN had better cross-system generalization capability. Compared with the other models, the ANN achieved an overall RMSE of 3.9 MPa on the test set. Subsequently, the ANN was selected as the surrogate model and coupled with a genetic algorithm for the inverse design of low-carbon clinkers. When the CO₂-emission reduction constraint was tightened to 25%, the genetic algorithm further favored low-carbon phases and generated candidate clinkers dominated by aluminate phases or the C4A3$ phase, with CO2 emissions of approximately 555 kg CO₂/t. Experimental validation showed that the measured 28 d compressive strengths of the candidate clinker compositions were close to the model predictions. The prediction errors for all specimens were lower than 10% of the measured strengths.

This study constructs a clinker dataset based on a HT preparation procedure and establishes an ML prediction and constrained-optimization framework for screening low-carbon clinkers. The results show that the dataset is characterized by broad compositional coverage, local sparsity, and pronounced system heterogeneity. Therefore, the model needs to balance nonlinear representation capability with regularization constraints. Comparisons on the validation sets and the test set show that the low-capacity ANN with moderate regularization achieves the best performance in terms of prediction accuracy. Its RMSE is 3.9 MPa on the overall test set, and the RMSE values for the in-system and out-of-system subsets are 3.1 MPa and 4.2 MPa, respectively, indicating that it can serve as the surrogate model for inverse design of low-carbon clinker. XGBoost can be regarded as a competitive potential surrogate model, but it tends to produce relatively large prediction errors for some specimens. By coupling the ANN surrogate model with a genetic algorithm, four candidate low-carbon clinkers satisfying the 28 d compressive strength and CO₂ emission constraints are screened. Experimental validation confirms good agreement between the predicted and measured compressive strengths. These results demonstrate that the proposed dataset construction and surrogate model screening strategy can be used to explore of novel low-carbon clinkers.