Resolving the Data Dilemma in Catalyst Design: A Generative-Enhanced Manifold-Preserving Machine Learning Method for CO2-to-Aromatics

Direct CO2 hydrogenation to aromatics (CTA) offers a sustainable alternative to petroleum-based aromatic production, yet its catalyst design is severely hindered by data scarcity, feature heterogeneity, and conflicting multiobjective requirements. Therefore, this study proposes a generative enhancement and manifold preservation-driven interpretable machine learning (GEMP-ML) framework. It integrates heuristic hybrid encoding, hierarchical missing value management, generative data augmentation, and active-learning-based sample screening to construct an information-enriched catalyst training set. To enable semantically faithful numerical representation of high-cardinality categorical features, a best-first sampling strategy is designed to identify the optimal hybrid encoding scheme from 89 candidate combinations specifically for these categorical variables. Combined with hierarchical semantic missing-value management and high-fidelity imputation algorithms, the data integrity of the CTA dataset is restored while preserving the semantic integrity of categorical features. GMM-based generative augmentation combined with diversity-aware active learning improved the utilization of synthetic data, and the CatBoost surrogate model achieved an average test set R2 of 0.973 across the three targets. Complementary interpretability analysis further identifies key descriptors governing catalyst performance, guiding the Pareto optimization of catalyst formulations and operating parameters. The recommended Zn–Al/HZSM-5 catalyst is predicted to achieve approximately 36.5% CO2 conversion, 78.6% C5+ selectivity, and 0.35% CO selectivity. The model is further validated against two independent experimental datasets covering 100 h time-on-stream, with prediction errors of <8%, supporting its ability to reproduce the observed performance trends under the evaluated conditions. This generalizable framework provides a systematic, data-driven paradigm for accelerating the design of high-performance catalysts in data-sparse catalytic systems.