Although the trial-and-error methodology remains predominant in polymer design, it is time-consuming and costly. Faced with these limitations, this thesis falls within the emerging field of polymer informatics by integrating artificial intelligence methods and domain expertise. It is organized around two main themes: on one hand, a meta-analysis based on knowledge modeling; on the other hand, the study of structure-property relationships for the rational design of materials, particularly epoxy–amine systems and polyimides.

The study is structured around four achievements.
First, we introduce CCASL (Counterexamples to Comparative Analysis of Scientific Literature), a methodology for automatically detecting and analyzing contradictions in polymers scientific publications. By leveraging relaxed functional dependencies and large-scale table extraction, CCASL highlights typographical and methodological inconsistencies, facilitating scientific validation and knowledge refinement.

Second, we propose a dual-embedding framework to accurately predict polymer glass transition temperature (T_g), a key property in determining application areas. This model combines a standard embedding with a property-aware embedding fine-tuned on T_g similarity, improving overall prediction accuracy.

Third, we investigate the influence of score aggregation functions in multi-score reinforcement learning for high T_g polyimides generation. Several strategies, including the geometric mean (GeoMean) and our proposed ExpAgg function, are compared to maximize the diversity and performance of generated polyimides.

Fourth, we introduce a synthesizability score (Syn-Score) based on molecular symmetry and sulfonyl group presence, guiding the generator toward experimentally feasible polyimides structures.

Overall, this thesis demonstrates how the combination of domain-specific chemical knowledge and artificial intelligence tools can accelerate the discovery of polymeric materials by coherently integrating modeling, prediction, generation, and experimental knowledge.