The AMETHYST-T2[1] project aims to revolutionize polymer design by shifting from the traditional trial- and-error approach to an AI-driven one. Currently, chemists largely rely on their knowledge and intuition, making sometimes the outcomes of polymer properties unpredictable. To boost the discovery of new polymers or new formulations, the purpose of this project is to develop an AI system capable of understanding the relation between characteristics and properties. Previous initiatives like the Polymer Genome project tried to propose polymers for target properties [2]. Dealing this topic requires interdisciplinarity considering individual constraints. In the area of machine learning, one indisputable truth reigns: data is the lifeblood that fuels its power. Delving into the captivating world of polymers, we find ourselves contemplating the heterogeneity and diversity of data in polymer universe, a puzzle that presents us with either suitable databases or, unfortunately, an enigmatic void. We chose a strategic and challenging sub-domain of polymers known as ”Epoxy/Amine”(EA) networks. In addition to its enormous significance in the industry, EA is strategic due to its strong correlation between formulation and characterization. Furthermore, a study on EA could be generalized to similar polymer sub-domains. Dealing with EA is challenging due to its data scarcity. As of today, there is no organized and open serious database available. How can we acquire and organize EA databases? To address this question, we realize that there are three main sources of polymer databases: (1) literature, (2) high-throughput characterization, and (3) simulation. The literature source (1) has the advantage that it can contain the results of the other two. And guess what ? It’s very often open source ! By focusing on the potential embedded in EA publications, the presence of structured information within tables is of interest. Hence, we hypothesized that prioritizing tables present in the publications could make the analysis of thousands of PDFs less cumbersome and potentially more efficient. Extracting tables is a more intricate task than it appears. Equipped with knowledge of OCR, images analysis, and machine learning, we have proceeded to extract, classify based on relevance, and normalize tables, from a sample database of 1611 EA publications. This process has yielded a collection of 507 CSV files containing EA data (compositions and properties) and represents a potential of 2535 lines of information, far greater than the largest database on EA&ML in the literature ‘54' [3]. Nonetheless, the pipeline remains in constant evolution, as a future data integration requires specific normalization and the extraction of additional information from the parent PDFs using NLP techniques.